Java tutorial
/* * Licensed to the Apache Software Foundation (ASF) under one * or more contributor license agreements. See the NOTICE file * distributed with this work for additional information * regarding copyright ownership. The ASF licenses this file * to you under the Apache License, Version 2.0 (the * "License"); you may not use this file except in compliance * with the License. You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, * software distributed under the License is distributed on an * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY * KIND, either express or implied. See the License for the * specific language governing permissions and limitations * under the License. */ package org.apache.sysml.lops.compile; import java.io.IOException; import java.util.ArrayList; import java.util.Arrays; import java.util.HashMap; import java.util.HashSet; import org.apache.commons.logging.Log; import org.apache.commons.logging.LogFactory; import org.apache.hadoop.io.DoubleWritable; import org.apache.hadoop.io.IntWritable; import org.apache.hadoop.mapred.SequenceFileInputFormat; import org.apache.sysml.api.DMLScript; import org.apache.sysml.api.DMLScript.RUNTIME_PLATFORM; import org.apache.sysml.conf.ConfigurationManager; import org.apache.sysml.conf.DMLConfig; import org.apache.sysml.hops.AggBinaryOp; import org.apache.sysml.hops.BinaryOp; import org.apache.sysml.hops.Hop.FileFormatTypes; import org.apache.sysml.hops.HopsException; import org.apache.sysml.hops.OptimizerUtils; import org.apache.sysml.lops.AppendM; import org.apache.sysml.lops.BinaryM; import org.apache.sysml.lops.CombineBinary; import org.apache.sysml.lops.Data; import org.apache.sysml.lops.Data.OperationTypes; import org.apache.sysml.lops.FunctionCallCP; import org.apache.sysml.lops.Lop; import org.apache.sysml.lops.Lop.Type; import org.apache.sysml.lops.LopProperties.ExecLocation; import org.apache.sysml.lops.LopProperties.ExecType; import org.apache.sysml.lops.LopsException; import org.apache.sysml.lops.MapMult; import org.apache.sysml.lops.OutputParameters; import org.apache.sysml.lops.OutputParameters.Format; import org.apache.sysml.lops.PMMJ; import org.apache.sysml.lops.ParameterizedBuiltin; import org.apache.sysml.lops.PickByCount; import org.apache.sysml.lops.SortKeys; import org.apache.sysml.lops.Unary; import org.apache.sysml.parser.DataExpression; import org.apache.sysml.parser.Expression; import org.apache.sysml.parser.Expression.DataType; import org.apache.sysml.parser.ParameterizedBuiltinFunctionExpression; import org.apache.sysml.parser.StatementBlock; import org.apache.sysml.runtime.DMLRuntimeException; import org.apache.sysml.runtime.controlprogram.parfor.ProgramConverter; import org.apache.sysml.runtime.controlprogram.parfor.util.IDSequence; import org.apache.sysml.runtime.instructions.CPInstructionParser; import org.apache.sysml.runtime.instructions.Instruction; import org.apache.sysml.runtime.instructions.Instruction.INSTRUCTION_TYPE; import org.apache.sysml.runtime.instructions.InstructionParser; import org.apache.sysml.runtime.instructions.MRJobInstruction; import org.apache.sysml.runtime.instructions.SPInstructionParser; import org.apache.sysml.runtime.instructions.cp.CPInstruction; import org.apache.sysml.runtime.instructions.cp.CPInstruction.CPINSTRUCTION_TYPE; import org.apache.sysml.runtime.instructions.cp.VariableCPInstruction; import org.apache.sysml.runtime.matrix.MatrixCharacteristics; import org.apache.sysml.runtime.matrix.data.InputInfo; import org.apache.sysml.runtime.matrix.data.OutputInfo; import org.apache.sysml.runtime.matrix.sort.PickFromCompactInputFormat; /** * * Class to maintain a DAG of lops and compile it into * runtime instructions, incl piggybacking into jobs. * * @param <N> the class parameter has no affect and is * only kept for documentation purposes. */ public class Dag<N extends Lop> { private static final Log LOG = LogFactory.getLog(Dag.class.getName()); private static final int CHILD_BREAKS_ALIGNMENT = 2; private static final int CHILD_DOES_NOT_BREAK_ALIGNMENT = 1; private static final int MRCHILD_NOT_FOUND = 0; private static final int MR_CHILD_FOUND_BREAKS_ALIGNMENT = 4; private static final int MR_CHILD_FOUND_DOES_NOT_BREAK_ALIGNMENT = 5; private static IDSequence job_id = null; private static IDSequence var_index = null; private int total_reducers = -1; private String scratch = ""; private String scratchFilePath = null; private double gmrMapperFootprint = 0; static { job_id = new IDSequence(); var_index = new IDSequence(); } // hash set for all nodes in dag private ArrayList<Lop> nodes = null; /* * Hashmap to translates the nodes in the DAG to a sequence of numbers * key: Lop ID * value: Sequence Number (0 ... |DAG|) * * This map is primarily used in performing DFS on the DAG, and subsequently in performing ancestor-descendant checks. */ private HashMap<Long, Integer> IDMap = null; private class NodeOutput { String fileName; String varName; OutputInfo outInfo; ArrayList<Instruction> preInstructions; // instructions added before a MR instruction ArrayList<Instruction> postInstructions; // instructions added after a MR instruction ArrayList<Instruction> lastInstructions; NodeOutput() { fileName = null; varName = null; outInfo = null; preInstructions = new ArrayList<Instruction>(); postInstructions = new ArrayList<Instruction>(); lastInstructions = new ArrayList<Instruction>(); } public String getFileName() { return fileName; } public void setFileName(String fileName) { this.fileName = fileName; } public String getVarName() { return varName; } public void setVarName(String varName) { this.varName = varName; } public OutputInfo getOutInfo() { return outInfo; } public void setOutInfo(OutputInfo outInfo) { this.outInfo = outInfo; } public ArrayList<Instruction> getPreInstructions() { return preInstructions; } public void addPreInstruction(Instruction inst) { preInstructions.add(inst); } public ArrayList<Instruction> getPostInstructions() { return postInstructions; } public void addPostInstruction(Instruction inst) { postInstructions.add(inst); } public ArrayList<Instruction> getLastInstructions() { return lastInstructions; } public void addLastInstruction(Instruction inst) { lastInstructions.add(inst); } } public Dag() { //allocate internal data structures nodes = new ArrayList<Lop>(); IDMap = new HashMap<Long, Integer>(); // get number of reducers from dml config total_reducers = ConfigurationManager.getNumReducers(); } /////// // filename handling private String getFilePath() { if (scratchFilePath == null) { scratchFilePath = scratch + Lop.FILE_SEPARATOR + Lop.PROCESS_PREFIX + DMLScript.getUUID() + Lop.FILE_SEPARATOR + Lop.FILE_SEPARATOR + ProgramConverter.CP_ROOT_THREAD_ID + Lop.FILE_SEPARATOR; } return scratchFilePath; } public String getNextUniqueFilename() { return getFilePath() + "temp" + job_id.getNextID(); } public static String getNextUniqueVarname(DataType dt) { return (dt == DataType.MATRIX ? Lop.MATRIX_VAR_NAME_PREFIX : Lop.FRAME_VAR_NAME_PREFIX) + var_index.getNextID(); } /////// // Dag modifications /** * Method to add a node to the DAG. * * @param node low-level operator * @return true if node was not already present, false if not. */ public boolean addNode(Lop node) { if (nodes.contains(node)) return false; nodes.add(node); return true; } /** * Method to compile a dag generically * * @param sb statement block * @param config dml configuration * @return list of instructions * @throws LopsException if LopsException occurs * @throws IOException if IOException occurs * @throws DMLRuntimeException if DMLRuntimeException occurs */ public ArrayList<Instruction> getJobs(StatementBlock sb, DMLConfig config) throws LopsException, IOException, DMLRuntimeException { if (config != null) { total_reducers = config.getIntValue(DMLConfig.NUM_REDUCERS); scratch = config.getTextValue(DMLConfig.SCRATCH_SPACE) + "/"; } // hold all nodes in a vector (needed for ordering) ArrayList<Lop> node_v = new ArrayList<Lop>(); node_v.addAll(nodes); /* * Sort the nodes by topological order. * * 1) All nodes with level i appear prior to the nodes in level i+1. * 2) All nodes within a level are ordered by their ID i.e., in the order * they are created */ doTopologicalSort_strict_order(node_v); // do greedy grouping of operations ArrayList<Instruction> inst = doGreedyGrouping(sb, node_v); return inst; } private static void deleteUpdatedTransientReadVariables(StatementBlock sb, ArrayList<Lop> nodeV, ArrayList<Instruction> inst) throws DMLRuntimeException { if (sb == null) return; if (LOG.isTraceEnabled()) LOG.trace("In delete updated variables"); // CANDIDATE list of variables which could have been updated in this statement block HashMap<String, Lop> labelNodeMapping = new HashMap<String, Lop>(); // ACTUAL list of variables whose value is updated, AND the old value of the variable // is no longer accessible/used. HashSet<String> updatedLabels = new HashSet<String>(); HashMap<String, Lop> updatedLabelsLineNum = new HashMap<String, Lop>(); // first capture all transient read variables for (Lop node : nodeV) { if (node.getExecLocation() == ExecLocation.Data && ((Data) node).isTransient() && ((Data) node).getOperationType() == OperationTypes.READ && ((Data) node).getDataType() == DataType.MATRIX) { // "node" is considered as updated ONLY IF the old value is not used any more // So, make sure that this READ node does not feed into any (transient/persistent) WRITE boolean hasWriteParent = false; for (Lop p : node.getOutputs()) { if (p.getExecLocation() == ExecLocation.Data) { // if the "p" is of type Data, then it has to be a WRITE hasWriteParent = true; break; } } if (!hasWriteParent) { // node has no parent of type WRITE, so this is a CANDIDATE variable // add it to labelNodeMapping so that it is considered in further processing labelNodeMapping.put(node.getOutputParameters().getLabel(), node); } } } // capture updated transient write variables for (Lop node : nodeV) { if (node.getExecLocation() == ExecLocation.Data && ((Data) node).isTransient() && ((Data) node).getOperationType() == OperationTypes.WRITE && ((Data) node).getDataType() == DataType.MATRIX && labelNodeMapping.containsKey(node.getOutputParameters().getLabel()) // check to make sure corresponding (i.e., with the same label/name) transient read is present && !labelNodeMapping.containsValue(node.getInputs().get(0)) // check to avoid cases where transient read feeds into a transient write ) { updatedLabels.add(node.getOutputParameters().getLabel()); updatedLabelsLineNum.put(node.getOutputParameters().getLabel(), node); } } // generate RM instructions Instruction rm_inst = null; for (String label : updatedLabels) { rm_inst = VariableCPInstruction.prepareRemoveInstruction(label); rm_inst.setLocation(updatedLabelsLineNum.get(label)); if (LOG.isTraceEnabled()) LOG.trace(rm_inst.toString()); inst.add(rm_inst); } } private static void generateRemoveInstructions(StatementBlock sb, ArrayList<Instruction> deleteInst) throws DMLRuntimeException { if (sb == null) return; if (LOG.isTraceEnabled()) LOG.trace("In generateRemoveInstructions()"); Instruction inst = null; // RULE 1: if in IN and not in OUT, then there should be an rmvar or rmfilevar inst // (currently required for specific cases of external functions) for (String varName : sb.liveIn().getVariableNames()) { if (!sb.liveOut().containsVariable(varName)) { // DataType dt = in.getVariable(varName).getDataType(); // if( !(dt==DataType.MATRIX || dt==DataType.UNKNOWN) ) // continue; //skip rm instructions for non-matrix objects inst = VariableCPInstruction.prepareRemoveInstruction(varName); inst.setLocation(sb.getEndLine(), sb.getEndLine(), -1, -1); deleteInst.add(inst); if (LOG.isTraceEnabled()) LOG.trace(" Adding " + inst.toString()); } } // RULE 2: if in KILL and not in IN and not in OUT, then there should be an rmvar or rmfilevar inst // (currently required for specific cases of nested loops) // i.e., local variables which are created within the block, and used entirely within the block /*for (String varName : sb.getKill().getVariableNames()) { if ((!sb.liveIn().containsVariable(varName)) && (!sb.liveOut().containsVariable(varName))) { // DataType dt = // sb.getKill().getVariable(varName).getDataType(); // if( !(dt==DataType.MATRIX || dt==DataType.UNKNOWN) ) // continue; //skip rm instructions for non-matrix objects inst = createCleanupInstruction(varName); deleteInst.add(inst); if (DMLScript.DEBUG) System.out.println("Adding instruction (r2) " + inst.toString()); } }*/ } private static ArrayList<ArrayList<Lop>> createNodeVectors(int size) { ArrayList<ArrayList<Lop>> arr = new ArrayList<ArrayList<Lop>>(); // for each job type, we need to create a vector. // additionally, create another vector for execNodes for (int i = 0; i < size; i++) { arr.add(new ArrayList<Lop>()); } return arr; } private static void clearNodeVectors(ArrayList<ArrayList<Lop>> arr) { for (ArrayList<Lop> tmp : arr) { tmp.clear(); } } private static boolean isCompatible(ArrayList<Lop> nodes, JobType jt, int from, int to) throws LopsException { int base = jt.getBase(); for (Lop node : nodes) { if ((node.getCompatibleJobs() & base) == 0) { if (LOG.isTraceEnabled()) LOG.trace("Not compatible " + node.toString()); return false; } } return true; } /** * Function that determines if the two input nodes can be executed together * in at least one job. * * @param node1 low-level operator 1 * @param node2 low-level operator 2 * @return true if nodes can be executed together */ private static boolean isCompatible(Lop node1, Lop node2) { return ((node1.getCompatibleJobs() & node2.getCompatibleJobs()) > 0); } /** * Function that checks if the given node executes in the job specified by jt. * * @param node low-level operator * @param jt job type * @return true if node executes in the specified job type */ private static boolean isCompatible(Lop node, JobType jt) { if (jt == JobType.GMRCELL) jt = JobType.GMR; return ((node.getCompatibleJobs() & jt.getBase()) > 0); } /* * Add node, and its relevant children to job-specific node vectors. */ private void addNodeByJobType(Lop node, ArrayList<ArrayList<Lop>> arr, ArrayList<Lop> execNodes, boolean eliminate) throws LopsException { if (!eliminate) { // Check if this lop defines a MR job. if (node.definesMRJob()) { // find the corresponding JobType JobType jt = JobType.findJobTypeFromLop(node); if (jt == null) { throw new LopsException(node.printErrorLocation() + "No matching JobType is found for a the lop type: " + node.getType() + " \n"); } // Add "node" to corresponding job vector if (jt == JobType.GMR) { if (node.hasNonBlockedInputs()) { int gmrcell_index = JobType.GMRCELL.getId(); arr.get(gmrcell_index).add(node); int from = arr.get(gmrcell_index).size(); addChildren(node, arr.get(gmrcell_index), execNodes); int to = arr.get(gmrcell_index).size(); if (!isCompatible(arr.get(gmrcell_index), JobType.GMR, from, to)) // check against GMR only, not against GMRCELL throw new LopsException( node.printErrorLocation() + "Error during compatibility check \n"); } else { // if "node" (in this case, a group lop) has any inputs from RAND // then add it to RAND job. Otherwise, create a GMR job if (hasChildNode(node, arr.get(JobType.DATAGEN.getId()))) { arr.get(JobType.DATAGEN.getId()).add(node); // we should NOT call 'addChildren' because appropriate // child nodes would have got added to RAND job already } else { int gmr_index = JobType.GMR.getId(); arr.get(gmr_index).add(node); int from = arr.get(gmr_index).size(); addChildren(node, arr.get(gmr_index), execNodes); int to = arr.get(gmr_index).size(); if (!isCompatible(arr.get(gmr_index), JobType.GMR, from, to)) throw new LopsException( node.printErrorLocation() + "Error during compatibility check \n"); } } } else { int index = jt.getId(); arr.get(index).add(node); int from = arr.get(index).size(); addChildren(node, arr.get(index), execNodes); int to = arr.get(index).size(); // check if all added nodes are compatible with current job if (!isCompatible(arr.get(index), jt, from, to)) { throw new LopsException("Unexpected error in addNodeByType."); } } return; } } if (eliminate) { // Eliminated lops are directly added to GMR queue. // Note that eliminate flag is set only for 'group' lops if (node.hasNonBlockedInputs()) arr.get(JobType.GMRCELL.getId()).add(node); else arr.get(JobType.GMR.getId()).add(node); return; } /* * If this lop does not define a job, check if it uses the output of any * specialized job. i.e., if this lop has a child node in any of the * job-specific vector, then add it to the vector. Note: This lop must * be added to ONLY ONE of the job-specific vectors. */ int numAdded = 0; for (JobType j : JobType.values()) { if (j.getId() > 0 && hasDirectChildNode(node, arr.get(j.getId()))) { if (isCompatible(node, j)) { arr.get(j.getId()).add(node); numAdded += 1; } } } if (numAdded > 1) { throw new LopsException( "Unexpected error in addNodeByJobType(): A given lop can ONLY be added to a single job vector (numAdded = " + numAdded + ")."); } } /* * Remove the node from all job-specific node vectors. This method is * invoked from removeNodesForNextIteration(). */ private static void removeNodeByJobType(Lop node, ArrayList<ArrayList<Lop>> arr) { for (JobType jt : JobType.values()) if (jt.getId() > 0) arr.get(jt.getId()).remove(node); } /** * As some jobs only write one output, all operations in the mapper need to * be redone and cannot be marked as finished. * * @param execNodes list of exec low-level operators * @param jobNodes list of job low-level operators * @param finishedNodes list of finished low-level operators * @throws LopsException if LopsException occurs */ private void handleSingleOutputJobs(ArrayList<Lop> execNodes, ArrayList<ArrayList<Lop>> jobNodes, ArrayList<Lop> finishedNodes) throws LopsException { /* * If the input of a MMCJ/MMRJ job (must have executed in a Mapper) is used * by multiple lops then we should mark it as not-finished. */ ArrayList<Lop> nodesWithUnfinishedOutputs = new ArrayList<Lop>(); int[] jobIndices = { JobType.MMCJ.getId() }; Lop.Type[] lopTypes = { Lop.Type.MMCJ }; // TODO: SortByValue should be treated similar to MMCJ, since it can // only sort one file now for (int jobi = 0; jobi < jobIndices.length; jobi++) { int jindex = jobIndices[jobi]; if (!jobNodes.get(jindex).isEmpty()) { ArrayList<Lop> vec = jobNodes.get(jindex); // first find all nodes with more than one parent that is not finished. for (int i = 0; i < vec.size(); i++) { Lop node = vec.get(i); if (node.getExecLocation() == ExecLocation.MapOrReduce || node.getExecLocation() == ExecLocation.Map) { Lop MRparent = getParentNode(node, execNodes, ExecLocation.MapAndReduce); if (MRparent != null && MRparent.getType() == lopTypes[jobi]) { int numParents = node.getOutputs().size(); if (numParents > 1) { for (int j = 0; j < numParents; j++) { if (!finishedNodes.contains(node.getOutputs().get(j))) nodesWithUnfinishedOutputs.add(node); } } } } } // need to redo all nodes in nodesWithOutput as well as their children for (Lop node : vec) { if (node.getExecLocation() == ExecLocation.MapOrReduce || node.getExecLocation() == ExecLocation.Map) { if (nodesWithUnfinishedOutputs.contains(node)) finishedNodes.remove(node); if (hasParentNode(node, nodesWithUnfinishedOutputs)) finishedNodes.remove(node); } } } } } /** * Method to check if a lop can be eliminated from checking * * @param node low-level operator * @param execNodes list of exec nodes * @return true if lop can be eliminated */ private static boolean canEliminateLop(Lop node, ArrayList<Lop> execNodes) { // this function can only eliminate "aligner" lops such a group if (!node.isAligner()) return false; // find the child whose execLoc = 'MapAndReduce' int ret = getChildAlignment(node, execNodes, ExecLocation.MapAndReduce); if (ret == CHILD_BREAKS_ALIGNMENT) return false; else if (ret == CHILD_DOES_NOT_BREAK_ALIGNMENT) return true; else if (ret == MRCHILD_NOT_FOUND) return false; else if (ret == MR_CHILD_FOUND_BREAKS_ALIGNMENT) return false; else if (ret == MR_CHILD_FOUND_DOES_NOT_BREAK_ALIGNMENT) return true; else throw new RuntimeException("Should not happen. \n"); } /** * Method to generate createvar instructions, which creates a new entry * in the symbol table. One instruction is generated for every LOP that is * 1) type Data and * 2) persistent and * 3) matrix and * 4) read * * Transient reads needn't be considered here since the previous program * block would already create appropriate entries in the symbol table. * * @param nodes_v list of nodes * @param inst list of instructions * @throws LopsException if LopsException occurs * @throws IOException if IOException occurs */ private static void generateInstructionsForInputVariables(ArrayList<Lop> nodes_v, ArrayList<Instruction> inst) throws LopsException, IOException { for (Lop n : nodes_v) { if (n.getExecLocation() == ExecLocation.Data && !((Data) n).isTransient() && ((Data) n).getOperationType() == OperationTypes.READ && (n.getDataType() == DataType.MATRIX || n.getDataType() == DataType.FRAME)) { if (!((Data) n).isLiteral()) { try { String inst_string = n.getInstructions(); CPInstruction currInstr = CPInstructionParser.parseSingleInstruction(inst_string); currInstr.setLocation(n); inst.add(currInstr); } catch (DMLRuntimeException e) { throw new LopsException(n.printErrorLocation() + "error generating instructions from input variables in Dag -- \n", e); } } } } } /** * Determine whether to send <code>node</code> to MR or to process it in the control program. * It is sent to MR in the following cases: * * 1) if input lop gets processed in MR then <code>node</code> can be piggybacked * * 2) if the exectype of write lop itself is marked MR i.e., memory estimate > memory budget. * * @param node low-level operator * @return true if lop should be sent to MR */ private static boolean sendWriteLopToMR(Lop node) { if (DMLScript.rtplatform == RUNTIME_PLATFORM.SINGLE_NODE) return false; Lop in = node.getInputs().get(0); Format nodeFormat = node.getOutputParameters().getFormat(); // Case of a transient read feeding into only one output persistent binaryblock write // Move the temporary file on HDFS to required persistent location, insteadof copying. if (in.getExecLocation() == ExecLocation.Data && in.getOutputs().size() == 1 && !((Data) node).isTransient() && ((Data) in).isTransient() && ((Data) in).getOutputParameters().isBlocked() && node.getOutputParameters().isBlocked()) { return false; } //send write lop to MR if (1) it is marked with exec type MR (based on its memory estimate), or //(2) if the input lop is in MR and the write format allows to pack it into the same job (this does //not apply to csv write because MR csvwrite is a separate MR job type) if (node.getExecType() == ExecType.MR || (in.getExecType() == ExecType.MR && nodeFormat != Format.CSV)) return true; else return false; } /** * Computes the memory footprint required to execute <code>node</code> in the mapper. * It is used only for those nodes that use inputs from distributed cache. The returned * value is utilized in limiting the number of instructions piggybacked onto a single GMR mapper. * * @param node low-level operator * @return memory footprint */ private static double computeFootprintInMapper(Lop node) { // Memory limits must be checked only for nodes that use distributed cache if (!node.usesDistributedCache()) // default behavior return 0.0; OutputParameters in1dims = node.getInputs().get(0).getOutputParameters(); OutputParameters in2dims = node.getInputs().get(1).getOutputParameters(); double footprint = 0; if (node instanceof MapMult) { int dcInputIndex = node.distributedCacheInputIndex()[0]; footprint = AggBinaryOp.getMapmmMemEstimate(in1dims.getNumRows(), in1dims.getNumCols(), in1dims.getRowsInBlock(), in1dims.getColsInBlock(), in1dims.getNnz(), in2dims.getNumRows(), in2dims.getNumCols(), in2dims.getRowsInBlock(), in2dims.getColsInBlock(), in2dims.getNnz(), dcInputIndex, false); } else if (node instanceof PMMJ) { int dcInputIndex = node.distributedCacheInputIndex()[0]; footprint = AggBinaryOp.getMapmmMemEstimate(in1dims.getNumRows(), 1, in1dims.getRowsInBlock(), in1dims.getColsInBlock(), in1dims.getNnz(), in2dims.getNumRows(), in2dims.getNumCols(), in2dims.getRowsInBlock(), in2dims.getColsInBlock(), in2dims.getNnz(), dcInputIndex, true); } else if (node instanceof AppendM) { footprint = BinaryOp.footprintInMapper(in1dims.getNumRows(), in1dims.getNumCols(), in2dims.getNumRows(), in2dims.getNumCols(), in1dims.getRowsInBlock(), in1dims.getColsInBlock()); } else if (node instanceof BinaryM) { footprint = BinaryOp.footprintInMapper(in1dims.getNumRows(), in1dims.getNumCols(), in2dims.getNumRows(), in2dims.getNumCols(), in1dims.getRowsInBlock(), in1dims.getColsInBlock()); } else { // default behavior return 0.0; } return footprint; } /** * Determines if <code>node</code> can be executed in current round of MR jobs or if it needs to be queued for later rounds. * If the total estimated footprint (<code>node</code> and previously added nodes in GMR) is less than available memory on * the mappers then <code>node</code> can be executed in current round, and <code>true</code> is returned. Otherwise, * <code>node</code> must be queued and <code>false</code> is returned. * * @param node low-level operator * @param footprintInMapper mapper footprint * @return true if node can be executed in current round of jobs */ private static boolean checkMemoryLimits(Lop node, double footprintInMapper) { boolean addNode = true; // Memory limits must be checked only for nodes that use distributed cache if (!node.usesDistributedCache()) // default behavior return addNode; double memBudget = Math.min(AggBinaryOp.MAPMULT_MEM_MULTIPLIER, BinaryOp.APPEND_MEM_MULTIPLIER) * OptimizerUtils.getRemoteMemBudgetMap(true); if (footprintInMapper <= memBudget) return addNode; else return !addNode; } /** * Method to group a vector of sorted lops. * * @param sb statement block * @param node_v list of low-level operators * @return list of instructions * @throws LopsException if LopsException occurs * @throws IOException if IOException occurs * @throws DMLRuntimeException if DMLRuntimeException occurs */ private ArrayList<Instruction> doGreedyGrouping(StatementBlock sb, ArrayList<Lop> node_v) throws LopsException, IOException, DMLRuntimeException { if (LOG.isTraceEnabled()) LOG.trace("Grouping DAG ============"); // nodes to be executed in current iteration ArrayList<Lop> execNodes = new ArrayList<Lop>(); // nodes that have already been processed ArrayList<Lop> finishedNodes = new ArrayList<Lop>(); // nodes that are queued for the following iteration ArrayList<Lop> queuedNodes = new ArrayList<Lop>(); ArrayList<ArrayList<Lop>> jobNodes = createNodeVectors(JobType.getNumJobTypes()); // list of instructions ArrayList<Instruction> inst = new ArrayList<Instruction>(); //ArrayList<Instruction> preWriteDeleteInst = new ArrayList<Instruction>(); ArrayList<Instruction> writeInst = new ArrayList<Instruction>(); ArrayList<Instruction> deleteInst = new ArrayList<Instruction>(); ArrayList<Instruction> endOfBlockInst = new ArrayList<Instruction>(); // remove files for transient reads that are updated. deleteUpdatedTransientReadVariables(sb, node_v, writeInst); generateRemoveInstructions(sb, endOfBlockInst); generateInstructionsForInputVariables(node_v, inst); boolean done = false; String indent = " "; while (!done) { if (LOG.isTraceEnabled()) LOG.trace("Grouping nodes in DAG"); execNodes.clear(); queuedNodes.clear(); clearNodeVectors(jobNodes); gmrMapperFootprint = 0; for (Lop node : node_v) { // finished nodes don't need to be processed if (finishedNodes.contains(node)) continue; if (LOG.isTraceEnabled()) LOG.trace("Processing node (" + node.getID() + ") " + node.toString() + " exec nodes size is " + execNodes.size()); //if node defines MR job, make sure it is compatible with all //its children nodes in execNodes if (node.definesMRJob() && !compatibleWithChildrenInExecNodes(execNodes, node)) { if (LOG.isTraceEnabled()) LOG.trace(indent + "Queueing node " + node.toString() + " (code 1)"); queuedNodes.add(node); removeNodesForNextIteration(node, finishedNodes, execNodes, queuedNodes, jobNodes); continue; } // if child is queued, this node will be processed in the later // iteration if (hasChildNode(node, queuedNodes)) { if (LOG.isTraceEnabled()) LOG.trace(indent + "Queueing node " + node.toString() + " (code 2)"); queuedNodes.add(node); // if node has more than two inputs, // remove children that will be needed in a future // iterations // may also have to remove parent nodes of these children removeNodesForNextIteration(node, finishedNodes, execNodes, queuedNodes, jobNodes); continue; } // if inputs come from different jobs, then queue if (node.getInputs().size() >= 2) { int jobid = Integer.MIN_VALUE; boolean queueit = false; for (int idx = 0; idx < node.getInputs().size(); idx++) { int input_jobid = jobType(node.getInputs().get(idx), jobNodes); if (input_jobid != -1) { if (jobid == Integer.MIN_VALUE) jobid = input_jobid; else if (jobid != input_jobid) { queueit = true; break; } } } if (queueit) { if (LOG.isTraceEnabled()) LOG.trace(indent + "Queueing node " + node.toString() + " (code 3)"); queuedNodes.add(node); removeNodesForNextIteration(node, finishedNodes, execNodes, queuedNodes, jobNodes); continue; } } // See if this lop can be eliminated // This check is for "aligner" lops (e.g., group) boolean eliminate = false; eliminate = canEliminateLop(node, execNodes); if (eliminate) { if (LOG.isTraceEnabled()) LOG.trace(indent + "Adding -" + node.toString()); execNodes.add(node); finishedNodes.add(node); addNodeByJobType(node, jobNodes, execNodes, eliminate); continue; } // If the node defines a MR Job then make sure none of its // children that defines a MR Job are present in execNodes if (node.definesMRJob()) { if (hasMRJobChildNode(node, execNodes)) { // "node" must NOT be queued when node=group and the child that defines job is Rand // this is because "group" can be pushed into the "Rand" job. if (!(node.getType() == Lop.Type.Grouping && checkDataGenAsChildNode(node, execNodes))) { if (LOG.isTraceEnabled()) LOG.trace(indent + "Queueing node " + node.toString() + " (code 4)"); queuedNodes.add(node); removeNodesForNextIteration(node, finishedNodes, execNodes, queuedNodes, jobNodes); continue; } } } // if "node" has more than one input, and has a descendant lop // in execNodes that is of type RecordReader // then all its inputs must be ancestors of RecordReader. If // not, queue "node" if (node.getInputs().size() > 1 && hasChildNode(node, execNodes, ExecLocation.RecordReader)) { // get the actual RecordReader lop Lop rr_node = getChildNode(node, execNodes, ExecLocation.RecordReader); // all inputs of "node" must be ancestors of rr_node boolean queue_it = false; for (Lop n : node.getInputs()) { // each input should be ancestor of RecordReader lop if (!n.equals(rr_node) && !isChild(rr_node, n, IDMap)) { queue_it = true; // i.e., "node" must be queued break; } } if (queue_it) { // queue node if (LOG.isTraceEnabled()) LOG.trace(indent + "Queueing -" + node.toString() + " (code 5)"); queuedNodes.add(node); // TODO: does this have to be modified to handle // recordreader lops? removeNodesForNextIteration(node, finishedNodes, execNodes, queuedNodes, jobNodes); continue; } else { // nothing here.. subsequent checks have to be performed // on "node" ; } } // data node, always add if child not queued // only write nodes are kept in execnodes if (node.getExecLocation() == ExecLocation.Data) { Data dnode = (Data) node; boolean dnode_queued = false; if (dnode.getOperationType() == OperationTypes.READ) { if (LOG.isTraceEnabled()) LOG.trace(indent + "Adding Data -" + node.toString()); // TODO: avoid readScalar instruction, and read it on-demand just like the way Matrices are read in control program if (node.getDataType() == DataType.SCALAR //TODO: LEO check the following condition is still needed && node.getOutputParameters().getFile_name() != null) { // this lop corresponds to reading a scalar from HDFS file // add it to execNodes so that "readScalar" instruction gets generated execNodes.add(node); // note: no need to add it to any job vector } } else if (dnode.getOperationType() == OperationTypes.WRITE) { // Skip the transient write <code>node</code> if the input is a // transient read with the same variable name. i.e., a dummy copy. // Hence, <code>node</code> can be avoided. // TODO: this case should ideally be handled in the language layer // prior to the construction of Hops Dag Lop input = dnode.getInputs().get(0); if (dnode.isTransient() && input.getExecLocation() == ExecLocation.Data && ((Data) input).isTransient() && dnode.getOutputParameters().getLabel() .equals(input.getOutputParameters().getLabel())) { // do nothing, <code>node</code> must not processed any further. ; } else if (execNodes.contains(input) && !isCompatible(node, input) && sendWriteLopToMR(node)) { // input is in execNodes but it is not compatible with write lop. So, queue the write lop. if (LOG.isTraceEnabled()) LOG.trace(indent + "Queueing -" + node.toString()); queuedNodes.add(node); dnode_queued = true; } else { if (LOG.isTraceEnabled()) LOG.trace(indent + "Adding Data -" + node.toString()); execNodes.add(node); if (sendWriteLopToMR(node)) { addNodeByJobType(node, jobNodes, execNodes, false); } } } if (!dnode_queued) finishedNodes.add(node); continue; } // map or reduce node, can always be piggybacked with parent if (node.getExecLocation() == ExecLocation.MapOrReduce) { if (LOG.isTraceEnabled()) LOG.trace(indent + "Adding -" + node.toString()); execNodes.add(node); finishedNodes.add(node); addNodeByJobType(node, jobNodes, execNodes, false); continue; } // RecordReader node, add, if no parent needs reduce, else queue if (node.getExecLocation() == ExecLocation.RecordReader) { // "node" should not have any children in // execNodes .. it has to be the first one in the job! if (!hasChildNode(node, execNodes, ExecLocation.Map) && !hasChildNode(node, execNodes, ExecLocation.MapAndReduce)) { if (LOG.isTraceEnabled()) LOG.trace(indent + "Adding -" + node.toString()); execNodes.add(node); finishedNodes.add(node); addNodeByJobType(node, jobNodes, execNodes, false); } else { if (LOG.isTraceEnabled()) LOG.trace(indent + "Queueing -" + node.toString() + " (code 6)"); queuedNodes.add(node); removeNodesForNextIteration(node, finishedNodes, execNodes, queuedNodes, jobNodes); } continue; } // map node, add, if no parent needs reduce, else queue if (node.getExecLocation() == ExecLocation.Map) { boolean queueThisNode = false; int subcode = -1; if (node.usesDistributedCache()) { // if an input to <code>node</code> comes from distributed cache // then that input must get executed in one of the previous jobs. int[] dcInputIndexes = node.distributedCacheInputIndex(); for (int dcInputIndex : dcInputIndexes) { Lop dcInput = node.getInputs().get(dcInputIndex - 1); if ((dcInput.getType() != Lop.Type.Data && dcInput.getExecType() == ExecType.MR) && execNodes.contains(dcInput)) { queueThisNode = true; subcode = 1; } } // Limit the number of distributed cache inputs based on the available memory in mappers double memsize = computeFootprintInMapper(node); //gmrMapperFootprint += computeFootprintInMapper(node); if (gmrMapperFootprint > 0 && !checkMemoryLimits(node, gmrMapperFootprint + memsize)) { queueThisNode = true; subcode = 2; } if (!queueThisNode) gmrMapperFootprint += memsize; } if (!queueThisNode && !hasChildNode(node, execNodes, ExecLocation.MapAndReduce) && !hasMRJobChildNode(node, execNodes)) { if (LOG.isTraceEnabled()) LOG.trace(indent + "Adding -" + node.toString()); execNodes.add(node); finishedNodes.add(node); addNodeByJobType(node, jobNodes, execNodes, false); } else { if (LOG.isTraceEnabled()) LOG.trace(indent + "Queueing -" + node.toString() + " (code 7 - " + "subcode " + subcode + ")"); queuedNodes.add(node); removeNodesForNextIteration(node, finishedNodes, execNodes, queuedNodes, jobNodes); } continue; } // reduce node, make sure no parent needs reduce, else queue if (node.getExecLocation() == ExecLocation.MapAndReduce) { // TODO: statiko -- keep the middle condition // discuss about having a lop that is MapAndReduce but does // not define a job if (LOG.isTraceEnabled()) LOG.trace(indent + "Adding -" + node.toString()); execNodes.add(node); finishedNodes.add(node); addNodeByJobType(node, jobNodes, execNodes, eliminate); continue; } // aligned reduce, make sure a parent that is reduce exists if (node.getExecLocation() == ExecLocation.Reduce) { if (compatibleWithChildrenInExecNodes(execNodes, node) && (hasChildNode(node, execNodes, ExecLocation.MapAndReduce) || hasChildNode(node, execNodes, ExecLocation.Map))) { if (LOG.isTraceEnabled()) LOG.trace(indent + "Adding -" + node.toString()); execNodes.add(node); finishedNodes.add(node); addNodeByJobType(node, jobNodes, execNodes, false); } else { if (LOG.isTraceEnabled()) LOG.trace(indent + "Queueing -" + node.toString() + " (code 8)"); queuedNodes.add(node); removeNodesForNextIteration(node, finishedNodes, execNodes, queuedNodes, jobNodes); } continue; } // add Scalar to execNodes if it has no child in exec nodes // that will be executed in a MR job. if (node.getExecLocation() == ExecLocation.ControlProgram) { for (Lop lop : node.getInputs()) { if (execNodes.contains(lop) && !(lop.getExecLocation() == ExecLocation.Data) && !(lop.getExecLocation() == ExecLocation.ControlProgram)) { if (LOG.isTraceEnabled()) LOG.trace(indent + "Queueing -" + node.toString() + " (code 9)"); queuedNodes.add(node); removeNodesForNextIteration(node, finishedNodes, execNodes, queuedNodes, jobNodes); break; } } if (queuedNodes.contains(node)) continue; if (LOG.isTraceEnabled()) LOG.trace(indent + "Adding - scalar" + node.toString()); execNodes.add(node); addNodeByJobType(node, jobNodes, execNodes, false); finishedNodes.add(node); continue; } } // no work to do if (execNodes.isEmpty()) { if (!queuedNodes.isEmpty()) { //System.err.println("Queued nodes should be 0"); throw new LopsException("Queued nodes should not be 0 at this point \n"); } if (LOG.isTraceEnabled()) LOG.trace("All done! queuedNodes = " + queuedNodes.size()); done = true; } else { // work to do if (LOG.isTraceEnabled()) LOG.trace("Generating jobs for group -- Node count=" + execNodes.size()); // first process scalar instructions generateControlProgramJobs(execNodes, inst, writeInst, deleteInst); // copy unassigned lops in execnodes to gmrnodes for (int i = 0; i < execNodes.size(); i++) { Lop node = execNodes.get(i); if (jobType(node, jobNodes) == -1) { if (isCompatible(node, JobType.GMR)) { if (node.hasNonBlockedInputs()) { jobNodes.get(JobType.GMRCELL.getId()).add(node); addChildren(node, jobNodes.get(JobType.GMRCELL.getId()), execNodes); } else { jobNodes.get(JobType.GMR.getId()).add(node); addChildren(node, jobNodes.get(JobType.GMR.getId()), execNodes); } } else { if (LOG.isTraceEnabled()) LOG.trace(indent + "Queueing -" + node.toString() + " (code 10)"); execNodes.remove(i); finishedNodes.remove(node); queuedNodes.add(node); removeNodesForNextIteration(node, finishedNodes, execNodes, queuedNodes, jobNodes); } } } // next generate MR instructions if (!execNodes.isEmpty()) generateMRJobs(execNodes, inst, writeInst, deleteInst, jobNodes); handleSingleOutputJobs(execNodes, jobNodes, finishedNodes); } } // add write and delete inst at the very end. //inst.addAll(preWriteDeleteInst); inst.addAll(writeInst); inst.addAll(deleteInst); inst.addAll(endOfBlockInst); return inst; } private boolean compatibleWithChildrenInExecNodes(ArrayList<Lop> execNodes, Lop node) { for (Lop tmpNode : execNodes) { // for lops that execute in control program, compatibleJobs property is set to LopProperties.INVALID // we should not consider such lops in this check if (isChild(tmpNode, node, IDMap) && tmpNode.getExecLocation() != ExecLocation.ControlProgram //&& tmpNode.getCompatibleJobs() != LopProperties.INVALID && (tmpNode.getCompatibleJobs() & node.getCompatibleJobs()) == 0) return false; } return true; } /** * Exclude rmvar instruction for varname from deleteInst, if exists * * @param varName variable name * @param deleteInst list of instructions */ private static void excludeRemoveInstruction(String varName, ArrayList<Instruction> deleteInst) { //for(Instruction inst : deleteInst) { for (int i = 0; i < deleteInst.size(); i++) { Instruction inst = deleteInst.get(i); if ((inst.getType() == INSTRUCTION_TYPE.CONTROL_PROGRAM || inst.getType() == INSTRUCTION_TYPE.SPARK) && ((CPInstruction) inst).getCPInstructionType() == CPINSTRUCTION_TYPE.Variable && ((VariableCPInstruction) inst).isRemoveVariable(varName)) { deleteInst.remove(i); } } } /** * Generate rmvar instructions for the inputs, if their consumer count becomes zero. * * @param node low-level operator * @param inst list of instructions * @param delteInst list of instructions * @throws DMLRuntimeException if DMLRuntimeException occurs */ private void processConsumersForInputs(Lop node, ArrayList<Instruction> inst, ArrayList<Instruction> delteInst) throws DMLRuntimeException { // reduce the consumer count for all input lops // if the count becomes zero, then then variable associated w/ input can be removed for (Lop in : node.getInputs()) { if (DMLScript.ENABLE_DEBUG_MODE) { processConsumers(in, inst, delteInst, node); } else { processConsumers(in, inst, delteInst, null); } } } private static void processConsumers(Lop node, ArrayList<Instruction> inst, ArrayList<Instruction> deleteInst, Lop locationInfo) throws DMLRuntimeException { // reduce the consumer count for all input lops // if the count becomes zero, then then variable associated w/ input can be removed if (node.removeConsumer() == 0) { if (node.getExecLocation() == ExecLocation.Data && ((Data) node).isLiteral()) { return; } String label = node.getOutputParameters().getLabel(); Instruction currInstr = VariableCPInstruction.prepareRemoveInstruction(label); if (locationInfo != null) currInstr.setLocation(locationInfo); else currInstr.setLocation(node); inst.add(currInstr); excludeRemoveInstruction(label, deleteInst); } } /** * Method to generate instructions that are executed in Control Program. At * this point, this DAG has no dependencies on the MR dag. ie. none of the * inputs are outputs of MR jobs * * @param execNodes list of low-level operators * @param inst list of instructions * @param writeInst list of write instructions * @param deleteInst list of delete instructions * @throws LopsException if LopsException occurs * @throws DMLRuntimeException if DMLRuntimeException occurs */ private void generateControlProgramJobs(ArrayList<Lop> execNodes, ArrayList<Instruction> inst, ArrayList<Instruction> writeInst, ArrayList<Instruction> deleteInst) throws LopsException, DMLRuntimeException { // nodes to be deleted from execnodes ArrayList<Lop> markedNodes = new ArrayList<Lop>(); // variable names to be deleted ArrayList<String> var_deletions = new ArrayList<String>(); HashMap<String, Lop> var_deletionsLineNum = new HashMap<String, Lop>(); boolean doRmVar = false; for (int i = 0; i < execNodes.size(); i++) { Lop node = execNodes.get(i); doRmVar = false; // mark input scalar read nodes for deletion // TODO: statiko -- check if this condition ever evaluated to TRUE if (node.getExecLocation() == ExecLocation.Data && ((Data) node).getOperationType() == Data.OperationTypes.READ && ((Data) node).getDataType() == DataType.SCALAR && node.getOutputParameters().getFile_name() == null) { markedNodes.add(node); continue; } // output scalar instructions and mark nodes for deletion if (node.getExecLocation() == ExecLocation.ControlProgram) { if (node.getDataType() == DataType.SCALAR) { // Output from lops with SCALAR data type must // go into Temporary Variables (Var0, Var1, etc.) NodeOutput out = setupNodeOutputs(node, ExecType.CP, false, false); inst.addAll(out.getPreInstructions()); // dummy deleteInst.addAll(out.getLastInstructions()); } else { // Output from lops with non-SCALAR data type must // go into Temporary Files (temp0, temp1, etc.) NodeOutput out = setupNodeOutputs(node, ExecType.CP, false, false); inst.addAll(out.getPreInstructions()); boolean hasTransientWriteParent = false; for (Lop parent : node.getOutputs()) { if (parent.getExecLocation() == ExecLocation.Data && ((Data) parent).getOperationType() == Data.OperationTypes.WRITE && ((Data) parent).isTransient()) { hasTransientWriteParent = true; break; } } if (!hasTransientWriteParent) { deleteInst.addAll(out.getLastInstructions()); } else { var_deletions.add(node.getOutputParameters().getLabel()); var_deletionsLineNum.put(node.getOutputParameters().getLabel(), node); } } String inst_string = ""; // Lops with arbitrary number of inputs (ParameterizedBuiltin, GroupedAggregate, DataGen) // are handled separately, by simply passing ONLY the output variable to getInstructions() if (node.getType() == Lop.Type.ParameterizedBuiltin || node.getType() == Lop.Type.GroupedAgg || node.getType() == Lop.Type.DataGen) { inst_string = node.getInstructions(node.getOutputParameters().getLabel()); } // Lops with arbitrary number of inputs and outputs are handled // separately as well by passing arrays of inputs and outputs else if (node.getType() == Lop.Type.FunctionCallCP) { String[] inputs = new String[node.getInputs().size()]; String[] outputs = new String[node.getOutputs().size()]; int count = 0; for (Lop in : node.getInputs()) inputs[count++] = in.getOutputParameters().getLabel(); count = 0; for (Lop out : node.getOutputs()) { outputs[count++] = out.getOutputParameters().getLabel(); } inst_string = node.getInstructions(inputs, outputs); } else if (node.getType() == Lop.Type.MULTIPLE_CP) { // ie, MultipleCP class inst_string = node.getInstructions(node.getOutputParameters().getLabel()); } else { if (node.getInputs().isEmpty()) { // currently, such a case exists only for Rand lop inst_string = node.getInstructions(node.getOutputParameters().getLabel()); } else if (node.getInputs().size() == 1) { inst_string = node.getInstructions(node.getInputs().get(0).getOutputParameters().getLabel(), node.getOutputParameters().getLabel()); } else if (node.getInputs().size() == 2) { inst_string = node.getInstructions(node.getInputs().get(0).getOutputParameters().getLabel(), node.getInputs().get(1).getOutputParameters().getLabel(), node.getOutputParameters().getLabel()); } else if (node.getInputs().size() == 3 || node.getType() == Type.Ternary) { inst_string = node.getInstructions(node.getInputs().get(0).getOutputParameters().getLabel(), node.getInputs().get(1).getOutputParameters().getLabel(), node.getInputs().get(2).getOutputParameters().getLabel(), node.getOutputParameters().getLabel()); } else if (node.getInputs().size() == 4) { inst_string = node.getInstructions(node.getInputs().get(0).getOutputParameters().getLabel(), node.getInputs().get(1).getOutputParameters().getLabel(), node.getInputs().get(2).getOutputParameters().getLabel(), node.getInputs().get(3).getOutputParameters().getLabel(), node.getOutputParameters().getLabel()); } else if (node.getInputs().size() == 5) { inst_string = node.getInstructions(node.getInputs().get(0).getOutputParameters().getLabel(), node.getInputs().get(1).getOutputParameters().getLabel(), node.getInputs().get(2).getOutputParameters().getLabel(), node.getInputs().get(3).getOutputParameters().getLabel(), node.getInputs().get(4).getOutputParameters().getLabel(), node.getOutputParameters().getLabel()); } else if (node.getInputs().size() == 6) { inst_string = node.getInstructions(node.getInputs().get(0).getOutputParameters().getLabel(), node.getInputs().get(1).getOutputParameters().getLabel(), node.getInputs().get(2).getOutputParameters().getLabel(), node.getInputs().get(3).getOutputParameters().getLabel(), node.getInputs().get(4).getOutputParameters().getLabel(), node.getInputs().get(5).getOutputParameters().getLabel(), node.getOutputParameters().getLabel()); } else if (node.getInputs().size() == 7) { inst_string = node.getInstructions(node.getInputs().get(0).getOutputParameters().getLabel(), node.getInputs().get(1).getOutputParameters().getLabel(), node.getInputs().get(2).getOutputParameters().getLabel(), node.getInputs().get(3).getOutputParameters().getLabel(), node.getInputs().get(4).getOutputParameters().getLabel(), node.getInputs().get(5).getOutputParameters().getLabel(), node.getInputs().get(6).getOutputParameters().getLabel(), node.getOutputParameters().getLabel()); } else if (node.getInputs().size() == 13) { // Used for im2col and reshape_col inst_string = node.getInstructions(node.getInputs().get(0).getOutputParameters().getLabel(), node.getInputs().get(1).getOutputParameters().getLabel(), node.getInputs().get(2).getOutputParameters().getLabel(), node.getInputs().get(3).getOutputParameters().getLabel(), node.getInputs().get(4).getOutputParameters().getLabel(), node.getInputs().get(5).getOutputParameters().getLabel(), node.getInputs().get(6).getOutputParameters().getLabel(), node.getInputs().get(7).getOutputParameters().getLabel(), node.getInputs().get(8).getOutputParameters().getLabel(), node.getInputs().get(9).getOutputParameters().getLabel(), node.getInputs().get(10).getOutputParameters().getLabel(), node.getInputs().get(11).getOutputParameters().getLabel(), node.getInputs().get(12).getOutputParameters().getLabel(), node.getOutputParameters().getLabel()); } else if (node.getInputs().size() == 14) { // Used for pooling_backward inst_string = node.getInstructions(node.getInputs().get(0).getOutputParameters().getLabel(), node.getInputs().get(1).getOutputParameters().getLabel(), node.getInputs().get(2).getOutputParameters().getLabel(), node.getInputs().get(3).getOutputParameters().getLabel(), node.getInputs().get(4).getOutputParameters().getLabel(), node.getInputs().get(5).getOutputParameters().getLabel(), node.getInputs().get(6).getOutputParameters().getLabel(), node.getInputs().get(7).getOutputParameters().getLabel(), node.getInputs().get(8).getOutputParameters().getLabel(), node.getInputs().get(9).getOutputParameters().getLabel(), node.getInputs().get(10).getOutputParameters().getLabel(), node.getInputs().get(11).getOutputParameters().getLabel(), node.getInputs().get(12).getOutputParameters().getLabel(), node.getInputs().get(13).getOutputParameters().getLabel(), node.getOutputParameters().getLabel()); } else if (node.getInputs().size() == 15) { // Used for fused conv2d_bias_add inst_string = node.getInstructions(node.getInputs().get(0).getOutputParameters().getLabel(), node.getInputs().get(1).getOutputParameters().getLabel(), node.getInputs().get(2).getOutputParameters().getLabel(), node.getInputs().get(3).getOutputParameters().getLabel(), node.getInputs().get(4).getOutputParameters().getLabel(), node.getInputs().get(5).getOutputParameters().getLabel(), node.getInputs().get(6).getOutputParameters().getLabel(), node.getInputs().get(7).getOutputParameters().getLabel(), node.getInputs().get(8).getOutputParameters().getLabel(), node.getInputs().get(9).getOutputParameters().getLabel(), node.getInputs().get(10).getOutputParameters().getLabel(), node.getInputs().get(11).getOutputParameters().getLabel(), node.getInputs().get(12).getOutputParameters().getLabel(), node.getInputs().get(13).getOutputParameters().getLabel(), node.getInputs().get(14).getOutputParameters().getLabel(), node.getOutputParameters().getLabel()); } else { throw new LopsException(node.printErrorLocation() + "Node with " + node.getInputs().size() + " inputs is not supported in CP yet! \n"); } } try { if (LOG.isTraceEnabled()) LOG.trace("Generating instruction - " + inst_string); Instruction currInstr = InstructionParser.parseSingleInstruction(inst_string); if (currInstr == null) { throw new LopsException("Error parsing the instruction:" + inst_string); } if (node._beginLine != 0) currInstr.setLocation(node); else if (!node.getOutputs().isEmpty()) currInstr.setLocation(node.getOutputs().get(0)); else if (!node.getInputs().isEmpty()) currInstr.setLocation(node.getInputs().get(0)); inst.add(currInstr); } catch (Exception e) { throw new LopsException( node.printErrorLocation() + "Problem generating simple inst - " + inst_string, e); } markedNodes.add(node); doRmVar = true; //continue; } else if (node.getExecLocation() == ExecLocation.Data) { Data dnode = (Data) node; Data.OperationTypes op = dnode.getOperationType(); if (op == Data.OperationTypes.WRITE) { NodeOutput out = null; if (sendWriteLopToMR(node)) { // In this case, Data WRITE lop goes into MR, and // we don't have to do anything here doRmVar = false; } else { out = setupNodeOutputs(node, ExecType.CP, false, false); if (dnode.getDataType() == DataType.SCALAR) { // processing is same for both transient and persistent scalar writes writeInst.addAll(out.getLastInstructions()); //inst.addAll(out.getLastInstructions()); doRmVar = false; } else { // setupNodeOutputs() handles both transient and persistent matrix writes if (dnode.isTransient()) { //inst.addAll(out.getPreInstructions()); // dummy ? deleteInst.addAll(out.getLastInstructions()); doRmVar = false; } else { // In case of persistent write lop, write instruction will be generated // and that instruction must be added to <code>inst</code> so that it gets // executed immediately. If it is added to <code>deleteInst</code> then it // gets executed at the end of program block's execution inst.addAll(out.getLastInstructions()); doRmVar = true; } } markedNodes.add(node); //continue; } } else { // generate a temp label to hold the value that is read from HDFS if (node.getDataType() == DataType.SCALAR) { node.getOutputParameters().setLabel(Lop.SCALAR_VAR_NAME_PREFIX + var_index.getNextID()); String io_inst = node.getInstructions(node.getOutputParameters().getLabel(), node.getOutputParameters().getFile_name()); CPInstruction currInstr = CPInstructionParser.parseSingleInstruction(io_inst); currInstr.setLocation(node); inst.add(currInstr); Instruction tempInstr = VariableCPInstruction .prepareRemoveInstruction(node.getOutputParameters().getLabel()); tempInstr.setLocation(node); deleteInst.add(tempInstr); } else { throw new LopsException("Matrix READs are not handled in CP yet!"); } markedNodes.add(node); doRmVar = true; //continue; } } // see if rmvar instructions can be generated for node's inputs if (doRmVar) processConsumersForInputs(node, inst, deleteInst); doRmVar = false; } for (String var : var_deletions) { Instruction rmInst = VariableCPInstruction.prepareRemoveInstruction(var); if (LOG.isTraceEnabled()) LOG.trace(" Adding var_deletions: " + rmInst.toString()); rmInst.setLocation(var_deletionsLineNum.get(var)); deleteInst.add(rmInst); } // delete all marked nodes for (Lop node : markedNodes) { execNodes.remove(node); } } /** * Method to remove all child nodes of a queued node that should be executed * in a following iteration. * * @param node low-level operator * @param finishedNodes list of finished nodes * @param execNodes list of exec nodes * @param queuedNodes list of queued nodes * @param jobvec list of lists of low-level operators * @throws LopsException if LopsException occurs */ private void removeNodesForNextIteration(Lop node, ArrayList<Lop> finishedNodes, ArrayList<Lop> execNodes, ArrayList<Lop> queuedNodes, ArrayList<ArrayList<Lop>> jobvec) throws LopsException { // only queued nodes with multiple inputs need to be handled. if (node.getInputs().size() == 1) return; //if all children are queued, then there is nothing to do. boolean allQueued = true; for (Lop input : node.getInputs()) { if (!queuedNodes.contains(input)) { allQueued = false; break; } } if (allQueued) return; if (LOG.isTraceEnabled()) LOG.trace(" Before remove nodes for next iteration -- size of execNodes " + execNodes.size()); // Determine if <code>node</code> has inputs from the same job or multiple jobs int jobid = Integer.MIN_VALUE; boolean inputs_in_same_job = true; for (Lop input : node.getInputs()) { int input_jobid = jobType(input, jobvec); if (jobid == Integer.MIN_VALUE) jobid = input_jobid; else if (jobid != input_jobid) { inputs_in_same_job = false; break; } } // Determine if there exist any unassigned inputs to <code>node</code> // Evaluate only those lops that execute in MR. boolean unassigned_inputs = false; for (Lop input : node.getInputs()) { //if ( input.getExecLocation() != ExecLocation.ControlProgram && jobType(input, jobvec) == -1 ) { if (input.getExecType() == ExecType.MR && !execNodes.contains(input)) { //jobType(input, jobvec) == -1 ) { unassigned_inputs = true; break; } } // Determine if any node's children are queued boolean child_queued = false; for (Lop input : node.getInputs()) { if (queuedNodes.contains(input)) { child_queued = true; break; } } if (LOG.isTraceEnabled()) { LOG.trace(" Property Flags:"); LOG.trace(" Inputs in same job: " + inputs_in_same_job); LOG.trace(" Unassigned inputs: " + unassigned_inputs); LOG.trace(" Child queued: " + child_queued); } // Evaluate each lop in <code>execNodes</code> for removal. // Add lops to be removed to <code>markedNodes</code>. ArrayList<Lop> markedNodes = new ArrayList<Lop>(); for (Lop tmpNode : execNodes) { if (LOG.isTraceEnabled()) { LOG.trace(" Checking for removal (" + tmpNode.getID() + ") " + tmpNode.toString()); } // if tmpNode is not a descendant of 'node', then there is no advantage in removing tmpNode for later iterations. if (!isChild(tmpNode, node, IDMap)) continue; // handle group input lops if (node.getInputs().contains(tmpNode) && tmpNode.isAligner()) { markedNodes.add(tmpNode); if (LOG.isTraceEnabled()) LOG.trace(" Removing for next iteration (code 1): (" + tmpNode.getID() + ") " + tmpNode.toString()); } //if (child_queued) { // if one of the children are queued, // remove some child nodes on other leg that may be needed later on. // For e.g. Group lop. if (!hasOtherQueuedParentNode(tmpNode, queuedNodes, node) && branchHasNoOtherUnExecutedParents(tmpNode, node, execNodes, finishedNodes)) { boolean queueit = false; int code = -1; switch (node.getExecLocation()) { case Map: if (branchCanBePiggyBackedMap(tmpNode, node, execNodes, queuedNodes, markedNodes)) queueit = true; code = 2; break; case MapAndReduce: if (branchCanBePiggyBackedMapAndReduce(tmpNode, node, execNodes, queuedNodes) && !tmpNode.definesMRJob()) queueit = true; code = 3; break; case Reduce: if (branchCanBePiggyBackedReduce(tmpNode, node, execNodes, queuedNodes)) queueit = true; code = 4; break; default: //do nothing } if (queueit) { if (LOG.isTraceEnabled()) LOG.trace(" Removing for next iteration (code " + code + "): (" + tmpNode.getID() + ") " + tmpNode.toString()); markedNodes.add(tmpNode); } } /* * "node" has no other queued children. * * If inputs are in the same job and "node" is of type * MapAndReduce, then remove nodes of all types other than * Reduce, MapAndReduce, and the ones that define a MR job as * they can be piggybacked later. * * e.g: A=Rand, B=Rand, C=A%*%B Here, both inputs of MMCJ lop * come from Rand job, and they should not be removed. * * Other examples: -- MMCJ whose children are of type * MapAndReduce (say GMR) -- Inputs coming from two different * jobs .. GMR & REBLOCK */ //boolean himr = hasOtherMapAndReduceParentNode(tmpNode, execNodes,node); //boolean bcbp = branchCanBePiggyBackedMapAndReduce(tmpNode, node, execNodes, finishedNodes); //System.out.println(" .. " + inputs_in_same_job + "," + himr + "," + bcbp); if ((inputs_in_same_job || unassigned_inputs) && node.getExecLocation() == ExecLocation.MapAndReduce && !hasOtherMapAndReduceParentNode(tmpNode, execNodes, node) // don't remove since it already piggybacked with a MapReduce node && branchCanBePiggyBackedMapAndReduce(tmpNode, node, execNodes, queuedNodes) && !tmpNode.definesMRJob()) { if (LOG.isTraceEnabled()) LOG.trace(" Removing for next iteration (code 5): (" + tmpNode.getID() + ") " + tmpNode.toString()); markedNodes.add(tmpNode); } } // for i // we also need to delete all parent nodes of marked nodes for (Lop enode : execNodes) { if (LOG.isTraceEnabled()) { LOG.trace(" Checking for removal - (" + enode.getID() + ") " + enode.toString()); } if (hasChildNode(enode, markedNodes) && !markedNodes.contains(enode)) { markedNodes.add(enode); if (LOG.isTraceEnabled()) LOG.trace( " Removing for next iteration (code 6) (" + enode.getID() + ") " + enode.toString()); } } if (execNodes.size() != markedNodes.size()) { // delete marked nodes from finishedNodes and execNodes // add to queued nodes for (Lop n : markedNodes) { if (n.usesDistributedCache()) gmrMapperFootprint -= computeFootprintInMapper(n); finishedNodes.remove(n); execNodes.remove(n); removeNodeByJobType(n, jobvec); queuedNodes.add(n); } } } private boolean branchCanBePiggyBackedReduce(Lop tmpNode, Lop node, ArrayList<Lop> execNodes, ArrayList<Lop> queuedNodes) { if (node.getExecLocation() != ExecLocation.Reduce) return false; // if tmpNode is descendant of any queued child of node, then branch can not be piggybacked for (Lop ni : node.getInputs()) { if (queuedNodes.contains(ni) && isChild(tmpNode, ni, IDMap)) return false; } for (Lop n : execNodes) { if (n.equals(node)) continue; if (n.equals(tmpNode) && n.getExecLocation() != ExecLocation.Map && n.getExecLocation() != ExecLocation.MapOrReduce) return false; // check if n is on the branch tmpNode->*->node if (isChild(n, node, IDMap) && isChild(tmpNode, n, IDMap)) { if (!node.getInputs().contains(tmpNode) // redundant && n.getExecLocation() != ExecLocation.Map && n.getExecLocation() != ExecLocation.MapOrReduce) return false; } } return true; } private boolean branchCanBePiggyBackedMap(Lop tmpNode, Lop node, ArrayList<Lop> execNodes, ArrayList<Lop> queuedNodes, ArrayList<Lop> markedNodes) { if (node.getExecLocation() != ExecLocation.Map) return false; // if tmpNode is descendant of any queued child of node, then branch can not be piggybacked for (Lop ni : node.getInputs()) { if (queuedNodes != null && queuedNodes.contains(ni) && isChild(tmpNode, ni, IDMap)) return false; } // since node.location=Map: only Map & MapOrReduce lops must be considered if (tmpNode.definesMRJob() || (tmpNode.getExecLocation() != ExecLocation.Map && tmpNode.getExecLocation() != ExecLocation.MapOrReduce)) return false; // if there exist a node "dcInput" that is // -- a) parent of tmpNode, and b) feeds into "node" via distributed cache // then, tmpNode should not be removed. // "dcInput" must be executed prior to "node", and removal of tmpNode does not make that happen. if (node.usesDistributedCache()) { for (int dcInputIndex : node.distributedCacheInputIndex()) { Lop dcInput = node.getInputs().get(dcInputIndex - 1); if (isChild(tmpNode, dcInput, IDMap)) return false; } } // if tmpNode requires an input from distributed cache, // remove tmpNode only if that input can fit into mappers' memory. If not, if (tmpNode.usesDistributedCache()) { double memsize = computeFootprintInMapper(tmpNode); if (node.usesDistributedCache()) memsize += computeFootprintInMapper(node); if (markedNodes != null) { for (Lop n : markedNodes) { if (n.usesDistributedCache()) memsize += computeFootprintInMapper(n); } } if (!checkMemoryLimits(node, memsize)) { return false; } } if ((tmpNode.getCompatibleJobs() & node.getCompatibleJobs()) > 0) return true; else return false; } /** * Function that checks if <code>tmpNode</code> can be piggybacked with MapAndReduce * lop <code>node</code>. * * Decision depends on the exec location of <code>tmpNode</code>. If the exec location is: * MapAndReduce: CAN NOT be piggybacked since it defines its own MR job * Reduce: CAN NOT be piggybacked since it must execute before <code>node</code> * Map or MapOrReduce: CAN be piggybacked ONLY IF it is comatible w/ <code>tmpNode</code> * * @param tmpNode temporary low-level operator * @param node low-level operator * @param execNodes list of exec nodes * @param queuedNodes list of queued nodes * @return true if tmpNode can be piggbacked on node */ private boolean branchCanBePiggyBackedMapAndReduce(Lop tmpNode, Lop node, ArrayList<Lop> execNodes, ArrayList<Lop> queuedNodes) { if (node.getExecLocation() != ExecLocation.MapAndReduce) return false; JobType jt = JobType.findJobTypeFromLop(node); for (Lop n : execNodes) { if (n.equals(node)) continue; // Evaluate only nodes on the branch between tmpNode->..->node if (n.equals(tmpNode) || (isChild(n, node, IDMap) && isChild(tmpNode, n, IDMap))) { if (hasOtherMapAndReduceParentNode(tmpNode, queuedNodes, node)) return false; ExecLocation el = n.getExecLocation(); if (el != ExecLocation.Map && el != ExecLocation.MapOrReduce) return false; else if (!isCompatible(n, jt)) return false; } } return true; } private boolean branchHasNoOtherUnExecutedParents(Lop tmpNode, Lop node, ArrayList<Lop> execNodes, ArrayList<Lop> finishedNodes) { //if tmpNode has more than one unfinished output, return false if (tmpNode.getOutputs().size() > 1) { int cnt = 0; for (Lop output : tmpNode.getOutputs()) if (!finishedNodes.contains(output)) cnt++; if (cnt != 1) return false; } //check to see if any node between node and tmpNode has more than one unfinished output for (Lop n : execNodes) { if (n.equals(node) || n.equals(tmpNode)) continue; if (isChild(n, node, IDMap) && isChild(tmpNode, n, IDMap)) { int cnt = 0; for (Lop output : n.getOutputs()) { if (!finishedNodes.contains(output)) cnt++; } if (cnt != 1) return false; } } return true; } /** * Method to return the job index for a lop. * * @param lops low-level operator * @param jobvec list of lists of low-level operators * @return job index for a low-level operator * @throws LopsException if LopsException occurs */ private static int jobType(Lop lops, ArrayList<ArrayList<Lop>> jobvec) throws LopsException { for (JobType jt : JobType.values()) { int i = jt.getId(); if (i > 0 && jobvec.get(i) != null && jobvec.get(i).contains(lops)) { return i; } } return -1; } /** * Method to see if there is a node of type MapAndReduce between tmpNode and node * in given node collection * * @param tmpNode temporary low-level operator * @param nodeList list of low-level operators * @param node low-level operator * @return true if MapAndReduce node between tmpNode and node in nodeList */ private boolean hasOtherMapAndReduceParentNode(Lop tmpNode, ArrayList<Lop> nodeList, Lop node) { if (tmpNode.getExecLocation() == ExecLocation.MapAndReduce) return true; for (Lop n : tmpNode.getOutputs()) { if (nodeList.contains(n) && isChild(n, node, IDMap)) { if (!n.equals(node) && n.getExecLocation() == ExecLocation.MapAndReduce) return true; else return hasOtherMapAndReduceParentNode(n, nodeList, node); } } return false; } /** * Method to check if there is a queued node that is a parent of both tmpNode and node * * @param tmpNode temporary low-level operator * @param queuedNodes list of queued nodes * @param node low-level operator * @return true if there is a queued node that is a parent of tmpNode and node */ private boolean hasOtherQueuedParentNode(Lop tmpNode, ArrayList<Lop> queuedNodes, Lop node) { if (queuedNodes.isEmpty()) return false; boolean[] nodeMarked = node.get_reachable(); boolean[] tmpMarked = tmpNode.get_reachable(); long nodeid = IDMap.get(node.getID()); long tmpid = IDMap.get(tmpNode.getID()); for (Lop qnode : queuedNodes) { int id = IDMap.get(qnode.getID()); if ((id != nodeid && nodeMarked[id]) && (id != tmpid && tmpMarked[id])) return true; } return false; } /** * Method to print the lops grouped by job type * * @param jobNodes list of lists of low-level operators * @throws DMLRuntimeException if DMLRuntimeException occurs */ private static void printJobNodes(ArrayList<ArrayList<Lop>> jobNodes) throws DMLRuntimeException { if (LOG.isTraceEnabled()) { for (JobType jt : JobType.values()) { int i = jt.getId(); if (i > 0 && jobNodes.get(i) != null && !jobNodes.get(i).isEmpty()) { LOG.trace(jt.getName() + " Job Nodes:"); for (int j = 0; j < jobNodes.get(i).size(); j++) { LOG.trace( " " + jobNodes.get(i).get(j).getID() + ") " + jobNodes.get(i).get(j).toString()); } } } } } /** * Method to check if there exists any lops with ExecLocation=RecordReader * * @param nodes list of low-level operators * @param loc exec location * @return true if there is a node with RecordReader exec location */ private static boolean hasANode(ArrayList<Lop> nodes, ExecLocation loc) { for (Lop n : nodes) { if (n.getExecLocation() == ExecLocation.RecordReader) return true; } return false; } private ArrayList<ArrayList<Lop>> splitGMRNodesByRecordReader(ArrayList<Lop> gmrnodes) { // obtain the list of record reader nodes ArrayList<Lop> rrnodes = new ArrayList<Lop>(); for (Lop gmrnode : gmrnodes) { if (gmrnode.getExecLocation() == ExecLocation.RecordReader) rrnodes.add(gmrnode); } // We allocate one extra vector to hold lops that do not depend on any // recordreader lops ArrayList<ArrayList<Lop>> splitGMR = createNodeVectors(rrnodes.size() + 1); // flags to indicate whether a lop has been added to one of the node vectors boolean[] flags = new boolean[gmrnodes.size()]; Arrays.fill(flags, false); // first, obtain all ancestors of recordreader lops for (int rrid = 0; rrid < rrnodes.size(); rrid++) { // prepare node list for i^th record reader lop // add record reader lop splitGMR.get(rrid).add(rrnodes.get(rrid)); for (int j = 0; j < gmrnodes.size(); j++) { if (rrnodes.get(rrid).equals(gmrnodes.get(j))) flags[j] = true; else if (isChild(rrnodes.get(rrid), gmrnodes.get(j), IDMap)) { splitGMR.get(rrid).add(gmrnodes.get(j)); flags[j] = true; } } } // add all remaining lops to a separate job int jobindex = rrnodes.size(); // the last node vector for (int i = 0; i < gmrnodes.size(); i++) { if (!flags[i]) { splitGMR.get(jobindex).add(gmrnodes.get(i)); flags[i] = true; } } return splitGMR; } /** * Method to generate hadoop jobs. Exec nodes can contains a mixture of node * types requiring different mr jobs. This method breaks the job into * sub-types and then invokes the appropriate method to generate * instructions. * * @param execNodes list of exec nodes * @param inst list of instructions * @param writeinst list of write instructions * @param deleteinst list of delete instructions * @param jobNodes list of list of low-level operators * @throws LopsException if LopsException occurs * @throws DMLRuntimeException if DMLRuntimeException occurs */ private void generateMRJobs(ArrayList<Lop> execNodes, ArrayList<Instruction> inst, ArrayList<Instruction> writeinst, ArrayList<Instruction> deleteinst, ArrayList<ArrayList<Lop>> jobNodes) throws LopsException, DMLRuntimeException { printJobNodes(jobNodes); ArrayList<Instruction> rmvarinst = new ArrayList<Instruction>(); for (JobType jt : JobType.values()) { // do nothing, if jt = INVALID or ANY if (jt == JobType.INVALID || jt == JobType.ANY) continue; int index = jt.getId(); // job id is used as an index into jobNodes ArrayList<Lop> currNodes = jobNodes.get(index); // generate MR job if (currNodes != null && !currNodes.isEmpty()) { if (LOG.isTraceEnabled()) LOG.trace("Generating " + jt.getName() + " job"); if (jt.allowsRecordReaderInstructions() && hasANode(jobNodes.get(index), ExecLocation.RecordReader)) { // split the nodes by recordReader lops ArrayList<ArrayList<Lop>> rrlist = splitGMRNodesByRecordReader(jobNodes.get(index)); for (int i = 0; i < rrlist.size(); i++) { generateMapReduceInstructions(rrlist.get(i), inst, writeinst, deleteinst, rmvarinst, jt); } } else if (jt.allowsSingleShuffleInstruction()) { // These jobs allow a single shuffle instruction. // We should split the nodes so that a separate job is produced for each shuffle instruction. Lop.Type splittingLopType = jt.getShuffleLopType(); ArrayList<Lop> nodesForASingleJob = new ArrayList<Lop>(); for (int i = 0; i < jobNodes.get(index).size(); i++) { if (jobNodes.get(index).get(i).getType() == splittingLopType) { nodesForASingleJob.clear(); // Add the lop that defines the split nodesForASingleJob.add(jobNodes.get(index).get(i)); /* * Add the splitting lop's children. This call is redundant when jt=SORT * because a sort job ALWAYS has a SINGLE lop in the entire job * i.e., there are no children to add when jt=SORT. */ addChildren(jobNodes.get(index).get(i), nodesForASingleJob, jobNodes.get(index)); if (jt.isCompatibleWithParentNodes()) { /* * If the splitting lop is compatible with parent nodes * then they must be added to the job. For example, MMRJ lop * may have a Data(Write) lop as its parent, which can be * executed along with MMRJ. */ addParents(jobNodes.get(index).get(i), nodesForASingleJob, jobNodes.get(index)); } generateMapReduceInstructions(nodesForASingleJob, inst, writeinst, deleteinst, rmvarinst, jt); } } } else { // the default case generateMapReduceInstructions(jobNodes.get(index), inst, writeinst, deleteinst, rmvarinst, jt); } } } inst.addAll(rmvarinst); } /** * Method to add all parents of "node" in exec_n to node_v. * * @param node low-level operator * @param node_v list of nodes * @param exec_n list of nodes */ private void addParents(Lop node, ArrayList<Lop> node_v, ArrayList<Lop> exec_n) { for (Lop enode : exec_n) { if (isChild(node, enode, IDMap)) { if (!node_v.contains(enode)) { if (LOG.isTraceEnabled()) LOG.trace("Adding parent - " + enode.toString()); node_v.add(enode); } } } } /** * Method to add all relevant data nodes for set of exec nodes. * * @param node low-level operator * @param node_v list of nodes * @param exec_n list of nodes */ private static void addChildren(Lop node, ArrayList<Lop> node_v, ArrayList<Lop> exec_n) { // add child in exec nodes that is not of type scalar if (exec_n.contains(node) && node.getExecLocation() != ExecLocation.ControlProgram) { if (!node_v.contains(node)) { node_v.add(node); if (LOG.isTraceEnabled()) LOG.trace(" Added child " + node.toString()); } } if (!exec_n.contains(node)) return; // recurse for (Lop n : node.getInputs()) { addChildren(n, node_v, exec_n); } } /** * Method that determines the output format for a given node. * * @param node low-level operator * @param cellModeOverride override mode * @return output info * @throws LopsException if LopsException occurs */ private static OutputInfo getOutputInfo(Lop node, boolean cellModeOverride) throws LopsException { if ((node.getDataType() == DataType.SCALAR && node.getExecType() == ExecType.CP) || node instanceof FunctionCallCP) return null; OutputInfo oinfo = null; OutputParameters oparams = node.getOutputParameters(); if (oparams.isBlocked()) { if (!cellModeOverride) oinfo = OutputInfo.BinaryBlockOutputInfo; else { // output format is overridden, for example, due to recordReaderInstructions in the job oinfo = OutputInfo.BinaryCellOutputInfo; // record decision of overriding in lop's outputParameters so that // subsequent jobs that use this lop's output know the correct format. // TODO: ideally, this should be done by having a member variable in Lop // which stores the outputInfo. try { oparams.setDimensions(oparams.getNumRows(), oparams.getNumCols(), -1, -1, oparams.getNnz(), oparams.getUpdateType()); } catch (HopsException e) { throw new LopsException(node.printErrorLocation() + "error in getOutputInfo in Dag ", e); } } } else { if (oparams.getFormat() == Format.TEXT || oparams.getFormat() == Format.MM) oinfo = OutputInfo.TextCellOutputInfo; else if (oparams.getFormat() == Format.CSV) { oinfo = OutputInfo.CSVOutputInfo; } else { oinfo = OutputInfo.BinaryCellOutputInfo; } } /* Instead of following hardcoding, one must get this information from Lops */ if (node.getType() == Type.SortKeys && node.getExecType() == ExecType.MR) { if (((SortKeys) node).getOpType() == SortKeys.OperationTypes.Indexes) oinfo = OutputInfo.BinaryBlockOutputInfo; else oinfo = OutputInfo.OutputInfoForSortOutput; } else if (node.getType() == Type.CombineBinary) { // Output format of CombineBinary (CB) depends on how the output is consumed CombineBinary combine = (CombineBinary) node; if (combine.getOperation() == org.apache.sysml.lops.CombineBinary.OperationTypes.PreSort) { oinfo = OutputInfo.OutputInfoForSortInput; } else if (combine.getOperation() == org.apache.sysml.lops.CombineBinary.OperationTypes.PreCentralMoment || combine.getOperation() == org.apache.sysml.lops.CombineBinary.OperationTypes.PreCovUnweighted || combine .getOperation() == org.apache.sysml.lops.CombineBinary.OperationTypes.PreGroupedAggUnweighted) { oinfo = OutputInfo.WeightedPairOutputInfo; } } else if (node.getType() == Type.CombineTernary) { oinfo = OutputInfo.WeightedPairOutputInfo; } else if (node.getType() == Type.CentralMoment || node.getType() == Type.CoVariance) { // CMMR always operate in "cell mode", // and the output is always in cell format oinfo = OutputInfo.BinaryCellOutputInfo; } return oinfo; } private String prepareAssignVarInstruction(Lop input, Lop node) { StringBuilder sb = new StringBuilder(); sb.append(ExecType.CP); sb.append(Lop.OPERAND_DELIMITOR); sb.append("assignvar"); sb.append(Lop.OPERAND_DELIMITOR); sb.append(input.prepScalarInputOperand(ExecType.CP)); sb.append(Lop.OPERAND_DELIMITOR); sb.append(node.prepOutputOperand()); return sb.toString(); } /** * Method to setup output filenames and outputInfos, and to generate related instructions * * @param node low-level operator * @param et exec type * @param cellModeOverride override mode * @param copyTWrite ? * @return node output * @throws DMLRuntimeException if DMLRuntimeException occurs * @throws LopsException if LopsException occurs */ private NodeOutput setupNodeOutputs(Lop node, ExecType et, boolean cellModeOverride, boolean copyTWrite) throws DMLRuntimeException, LopsException { OutputParameters oparams = node.getOutputParameters(); NodeOutput out = new NodeOutput(); node.setConsumerCount(node.getOutputs().size()); // Compute the output format for this node out.setOutInfo(getOutputInfo(node, cellModeOverride)); // If node is NOT of type Data then we must generate // a variable to hold the value produced by this node // note: functioncallcp requires no createvar, rmvar since // since outputs are explicitly specified if (node.getExecLocation() != ExecLocation.Data) { if (node.getDataType() == DataType.SCALAR) { oparams.setLabel(Lop.SCALAR_VAR_NAME_PREFIX + var_index.getNextID()); out.setVarName(oparams.getLabel()); Instruction currInstr = VariableCPInstruction.prepareRemoveInstruction(oparams.getLabel()); currInstr.setLocation(node); out.addLastInstruction(currInstr); } else if (node instanceof ParameterizedBuiltin && ((ParameterizedBuiltin) node) .getOp() == org.apache.sysml.lops.ParameterizedBuiltin.OperationTypes.TRANSFORM) { ParameterizedBuiltin pbi = (ParameterizedBuiltin) node; Lop input = pbi.getNamedInput(ParameterizedBuiltinFunctionExpression.TF_FN_PARAM_DATA); if (input.getDataType() == DataType.FRAME) { // Output of transform is in CSV format, which gets subsequently reblocked // TODO: change it to output binaryblock Data dataInput = (Data) input; oparams.setFile_name(getNextUniqueFilename()); oparams.setLabel(getNextUniqueVarname(DataType.MATRIX)); // generate an instruction that creates a symbol table entry for the new variable in CSV format Data delimLop = (Data) dataInput.getNamedInputLop(DataExpression.DELIM_DELIMITER, DataExpression.DEFAULT_DELIM_DELIMITER); Instruction createvarInst = VariableCPInstruction.prepareCreateVariableInstruction( oparams.getLabel(), oparams.getFile_name(), true, DataType.MATRIX, OutputInfo.outputInfoToString(OutputInfo.CSVOutputInfo), new MatrixCharacteristics(oparams.getNumRows(), oparams.getNumCols(), -1, -1, oparams.getNnz()), oparams.getUpdateType(), false, delimLop.getStringValue(), true); createvarInst.setLocation(node); out.addPreInstruction(createvarInst); // temp file as well as the variable has to be deleted at the end Instruction currInstr = VariableCPInstruction.prepareRemoveInstruction(oparams.getLabel()); currInstr.setLocation(node); out.addLastInstruction(currInstr); // finally, add the generated filename and variable name to the list of outputs out.setFileName(oparams.getFile_name()); out.setVarName(oparams.getLabel()); } else { throw new LopsException("Input to transform() has an invalid type: " + input.getDataType() + ", it must be FRAME."); } } else if (!(node instanceof FunctionCallCP)) //general case { // generate temporary filename and a variable name to hold the // output produced by "rootNode" oparams.setFile_name(getNextUniqueFilename()); oparams.setLabel(getNextUniqueVarname(node.getDataType())); // generate an instruction that creates a symbol table entry for the new variable //String createInst = prepareVariableInstruction("createvar", node); //out.addPreInstruction(CPInstructionParser.parseSingleInstruction(createInst)); int rpb = (int) oparams.getRowsInBlock(); int cpb = (int) oparams.getColsInBlock(); Instruction createvarInst = VariableCPInstruction.prepareCreateVariableInstruction( oparams.getLabel(), oparams.getFile_name(), true, node.getDataType(), OutputInfo.outputInfoToString(getOutputInfo(node, false)), new MatrixCharacteristics(oparams.getNumRows(), oparams.getNumCols(), rpb, cpb, oparams.getNnz()), oparams.getUpdateType()); createvarInst.setLocation(node); out.addPreInstruction(createvarInst); // temp file as well as the variable has to be deleted at the end Instruction currInstr = VariableCPInstruction.prepareRemoveInstruction(oparams.getLabel()); currInstr.setLocation(node); out.addLastInstruction(currInstr); // finally, add the generated filename and variable name to the list of outputs out.setFileName(oparams.getFile_name()); out.setVarName(oparams.getLabel()); } else { // If the function call is set with output lops (e.g., multi return builtin), // generate a createvar instruction for each function output FunctionCallCP fcall = (FunctionCallCP) node; if (fcall.getFunctionOutputs() != null) { for (Lop fnOut : fcall.getFunctionOutputs()) { OutputParameters fnOutParams = fnOut.getOutputParameters(); //OutputInfo oinfo = getOutputInfo((N)fnOut, false); Instruction createvarInst = VariableCPInstruction.prepareCreateVariableInstruction( fnOutParams.getLabel(), getFilePath() + fnOutParams.getLabel(), true, fnOut.getDataType(), OutputInfo.outputInfoToString(getOutputInfo(fnOut, false)), new MatrixCharacteristics(fnOutParams.getNumRows(), fnOutParams.getNumCols(), (int) fnOutParams.getRowsInBlock(), (int) fnOutParams.getColsInBlock(), fnOutParams.getNnz()), oparams.getUpdateType()); if (node._beginLine != 0) createvarInst.setLocation(node); else createvarInst.setLocation(fnOut); out.addPreInstruction(createvarInst); } } } } // rootNode is of type Data else { if (node.getDataType() == DataType.SCALAR) { // generate assignment operations for final and transient writes if (oparams.getFile_name() == null && !(node instanceof Data && ((Data) node).isPersistentWrite())) { String io_inst = prepareAssignVarInstruction(node.getInputs().get(0), node); CPInstruction currInstr = CPInstructionParser.parseSingleInstruction(io_inst); if (node._beginLine != 0) currInstr.setLocation(node); else if (!node.getInputs().isEmpty()) currInstr.setLocation(node.getInputs().get(0)); out.addLastInstruction(currInstr); } else { //CP PERSISTENT WRITE SCALARS Lop fname = ((Data) node).getNamedInputLop(DataExpression.IO_FILENAME); String io_inst = node.getInstructions(node.getInputs().get(0).getOutputParameters().getLabel(), fname.getOutputParameters().getLabel()); CPInstruction currInstr = CPInstructionParser.parseSingleInstruction(io_inst); if (node._beginLine != 0) currInstr.setLocation(node); else if (!node.getInputs().isEmpty()) currInstr.setLocation(node.getInputs().get(0)); out.addLastInstruction(currInstr); } } else { if (((Data) node).isTransient()) { if (et == ExecType.CP) { // If transient matrix write is in CP then its input MUST be executed in CP as well. // get variable and filename associated with the input String inputFileName = node.getInputs().get(0).getOutputParameters().getFile_name(); String inputVarName = node.getInputs().get(0).getOutputParameters().getLabel(); String constVarName = oparams.getLabel(); String constFileName = inputFileName + constVarName; /* * Symbol Table state must change as follows: * * FROM: * mvar1 -> temp21 * * TO: * mVar1 -> temp21 * tVarH -> temp21 */ Instruction currInstr = VariableCPInstruction.prepareCopyInstruction(inputVarName, constVarName); currInstr.setLocation(node); out.addLastInstruction(currInstr); out.setFileName(constFileName); } else { if (copyTWrite) { Instruction currInstr = VariableCPInstruction.prepareCopyInstruction( node.getInputs().get(0).getOutputParameters().getLabel(), oparams.getLabel()); currInstr.setLocation(node); out.addLastInstruction(currInstr); return out; } /* * Since the "rootNode" is a transient data node, we first need to generate a * temporary filename as well as a variable name to hold the <i>immediate</i> * output produced by "rootNode". These generated HDFS filename and the * variable name must be changed at the end of an iteration/program block * so that the subsequent iteration/program block can correctly access the * generated data. Therefore, we need to distinguish between the following: * * 1) Temporary file name & variable name: They hold the immediate output * produced by "rootNode". Both names are generated below. * * 2) Constant file name & variable name: They are constant across iterations. * Variable name is given by rootNode's label that is created in the upper layers. * File name is generated by concatenating "temporary file name" and "constant variable name". * * Temporary files must be moved to constant files at the end of the iteration/program block. */ // generate temporary filename & var name String tempVarName = oparams.getLabel() + "temp"; String tempFileName = getNextUniqueFilename(); //String createInst = prepareVariableInstruction("createvar", tempVarName, node.getDataType(), node.getValueType(), tempFileName, oparams, out.getOutInfo()); //out.addPreInstruction(CPInstructionParser.parseSingleInstruction(createInst)); int rpb = (int) oparams.getRowsInBlock(); int cpb = (int) oparams.getColsInBlock(); Instruction createvarInst = VariableCPInstruction.prepareCreateVariableInstruction( tempVarName, tempFileName, true, node.getDataType(), OutputInfo.outputInfoToString(out.getOutInfo()), new MatrixCharacteristics(oparams.getNumRows(), oparams.getNumCols(), rpb, cpb, oparams.getNnz()), oparams.getUpdateType()); createvarInst.setLocation(node); out.addPreInstruction(createvarInst); String constVarName = oparams.getLabel(); String constFileName = tempFileName + constVarName; oparams.setFile_name(getFilePath() + constFileName); /* * Since this is a node that denotes a transient read/write, we need to make sure * that the data computed for a given variable in a given iteration is passed on * to the next iteration. This is done by generating miscellaneous instructions * that gets executed at the end of the program block. * * The state of the symbol table must change * * FROM: * tVarA -> temp21tVarA (old copy of temp21) * tVarAtemp -> temp21 (new copy that should override the old copy) * * TO: * tVarA -> temp21tVarA */ // rename the temp variable to constant variable (e.g., cpvar tVarAtemp tVarA) /*Instruction currInstr = VariableCPInstruction.prepareCopyInstruction(tempVarName, constVarName); if(DMLScript.ENABLE_DEBUG_MODE) { currInstr.setLineNum(node._beginLine); } out.addLastInstruction(currInstr); Instruction tempInstr = VariableCPInstruction.prepareRemoveInstruction(tempVarName); if(DMLScript.ENABLE_DEBUG_MODE) { tempInstr.setLineNum(node._beginLine); } out.addLastInstruction(tempInstr);*/ // Generate a single mvvar instruction (e.g., mvvar tempA A) // instead of two instructions "cpvar tempA A" and "rmvar tempA" Instruction currInstr = VariableCPInstruction.prepareMoveInstruction(tempVarName, constVarName); currInstr.setLocation(node); out.addLastInstruction(currInstr); // finally, add the temporary filename and variable name to the list of outputs out.setFileName(tempFileName); out.setVarName(tempVarName); } } // rootNode is not a transient write. It is a persistent write. else { if (et == ExecType.MR) { //MR PERSISTENT WRITE // create a variable to hold the result produced by this "rootNode" oparams.setLabel("pVar" + var_index.getNextID()); //String createInst = prepareVariableInstruction("createvar", node); //out.addPreInstruction(CPInstructionParser.parseSingleInstruction(createInst)); int rpb = (int) oparams.getRowsInBlock(); int cpb = (int) oparams.getColsInBlock(); Lop fnameLop = ((Data) node).getNamedInputLop(DataExpression.IO_FILENAME); String fnameStr = (fnameLop instanceof Data && ((Data) fnameLop).isLiteral()) ? fnameLop.getOutputParameters().getLabel() : Lop.VARIABLE_NAME_PLACEHOLDER + fnameLop.getOutputParameters().getLabel() + Lop.VARIABLE_NAME_PLACEHOLDER; Instruction createvarInst; // for MatrixMarket format, the creatvar will output the result to a temporary file in textcell format // the CP write instruction (post instruction) after the MR instruction will merge the result into a single // part MM format file on hdfs. if (oparams.getFormat() == Format.CSV) { String tempFileName = getNextUniqueFilename(); String createInst = node.getInstructions(tempFileName); createvarInst = CPInstructionParser.parseSingleInstruction(createInst); //NOTE: no instruction patching because final write from cp instruction String writeInst = node.getInstructions(oparams.getLabel(), fnameLop.getOutputParameters().getLabel()); CPInstruction currInstr = CPInstructionParser.parseSingleInstruction(writeInst); currInstr.setLocation(node); out.addPostInstruction(currInstr); // remove the variable CPInstruction tempInstr = CPInstructionParser.parseSingleInstruction("CP" + Lop.OPERAND_DELIMITOR + "rmfilevar" + Lop.OPERAND_DELIMITOR + oparams.getLabel() + Lop.VALUETYPE_PREFIX + Expression.ValueType.UNKNOWN + Lop.OPERAND_DELIMITOR + "true" + Lop.VALUETYPE_PREFIX + "BOOLEAN"); tempInstr.setLocation(node); out.addLastInstruction(tempInstr); } else if (oparams.getFormat() == Format.MM) { createvarInst = VariableCPInstruction.prepareCreateVariableInstruction( oparams.getLabel(), getNextUniqueFilename(), false, node.getDataType(), OutputInfo.outputInfoToString(getOutputInfo(node, false)), new MatrixCharacteristics(oparams.getNumRows(), oparams.getNumCols(), rpb, cpb, oparams.getNnz()), oparams.getUpdateType()); //NOTE: no instruction patching because final write from cp instruction String writeInst = node.getInstructions(oparams.getLabel(), fnameLop.getOutputParameters().getLabel()); CPInstruction currInstr = CPInstructionParser.parseSingleInstruction(writeInst); currInstr.setLocation(node); out.addPostInstruction(currInstr); // remove the variable CPInstruction tempInstr = CPInstructionParser.parseSingleInstruction("CP" + Lop.OPERAND_DELIMITOR + "rmfilevar" + Lop.OPERAND_DELIMITOR + oparams.getLabel() + Lop.VALUETYPE_PREFIX + Expression.ValueType.UNKNOWN + Lop.OPERAND_DELIMITOR + "true" + Lop.VALUETYPE_PREFIX + "BOOLEAN"); tempInstr.setLocation(node); out.addLastInstruction(tempInstr); } else { createvarInst = VariableCPInstruction.prepareCreateVariableInstruction( oparams.getLabel(), fnameStr, false, node.getDataType(), OutputInfo.outputInfoToString(getOutputInfo(node, false)), new MatrixCharacteristics(oparams.getNumRows(), oparams.getNumCols(), rpb, cpb, oparams.getNnz()), oparams.getUpdateType()); // remove the variable CPInstruction currInstr = CPInstructionParser.parseSingleInstruction("CP" + Lop.OPERAND_DELIMITOR + "rmfilevar" + Lop.OPERAND_DELIMITOR + oparams.getLabel() + Lop.VALUETYPE_PREFIX + Expression.ValueType.UNKNOWN + Lop.OPERAND_DELIMITOR + "false" + Lop.VALUETYPE_PREFIX + "BOOLEAN"); currInstr.setLocation(node); out.addLastInstruction(currInstr); } createvarInst.setLocation(node); out.addPreInstruction(createvarInst); // finally, add the filename and variable name to the list of outputs out.setFileName(oparams.getFile_name()); out.setVarName(oparams.getLabel()); } else { //CP PERSISTENT WRITE // generate a write instruction that writes matrix to HDFS Lop fname = ((Data) node).getNamedInputLop(DataExpression.IO_FILENAME); Instruction currInstr = null; Lop inputLop = node.getInputs().get(0); // Case of a transient read feeding into only one output persistent binaryblock write // Move the temporary file on HDFS to required persistent location, insteadof copying. if (inputLop.getExecLocation() == ExecLocation.Data && inputLop.getOutputs().size() == 1 && ((Data) inputLop).isTransient() && ((Data) inputLop).getOutputParameters().isBlocked() && node.getOutputParameters().isBlocked()) { // transient read feeding into persistent write in blocked representation // simply, move the file //prepare filename (literal or variable in order to support dynamic write) String fnameStr = (fname instanceof Data && ((Data) fname).isLiteral()) ? fname.getOutputParameters().getLabel() : Lop.VARIABLE_NAME_PLACEHOLDER + fname.getOutputParameters().getLabel() + Lop.VARIABLE_NAME_PLACEHOLDER; currInstr = (CPInstruction) VariableCPInstruction.prepareMoveInstruction( inputLop.getOutputParameters().getLabel(), fnameStr, "binaryblock"); } else { String io_inst = node.getInstructions( node.getInputs().get(0).getOutputParameters().getLabel(), fname.getOutputParameters().getLabel()); if (node.getExecType() == ExecType.SPARK) // This will throw an exception if the exectype of hop is set incorrectly // Note: the exec type and exec location of lops needs to be set to SPARK and ControlProgram respectively currInstr = SPInstructionParser.parseSingleInstruction(io_inst); else currInstr = CPInstructionParser.parseSingleInstruction(io_inst); } if (!node.getInputs().isEmpty() && node.getInputs().get(0)._beginLine != 0) currInstr.setLocation(node.getInputs().get(0)); else currInstr.setLocation(node); out.addLastInstruction(currInstr); } } } } return out; } /** * Method to generate MapReduce job instructions from a given set of nodes. * * @param execNodes list of exec nodes * @param inst list of instructions * @param writeinst list of write instructions * @param deleteinst list of delete instructions * @param rmvarinst list of rmvar instructions * @param jt job type * @throws LopsException if LopsException occurs * @throws DMLRuntimeException if DMLRuntimeException occurs */ private void generateMapReduceInstructions(ArrayList<Lop> execNodes, ArrayList<Instruction> inst, ArrayList<Instruction> writeinst, ArrayList<Instruction> deleteinst, ArrayList<Instruction> rmvarinst, JobType jt) throws LopsException, DMLRuntimeException { ArrayList<Byte> resultIndices = new ArrayList<Byte>(); ArrayList<String> inputs = new ArrayList<String>(); ArrayList<String> outputs = new ArrayList<String>(); ArrayList<InputInfo> inputInfos = new ArrayList<InputInfo>(); ArrayList<OutputInfo> outputInfos = new ArrayList<OutputInfo>(); ArrayList<Long> numRows = new ArrayList<Long>(); ArrayList<Long> numCols = new ArrayList<Long>(); ArrayList<Long> numRowsPerBlock = new ArrayList<Long>(); ArrayList<Long> numColsPerBlock = new ArrayList<Long>(); ArrayList<String> mapperInstructions = new ArrayList<String>(); ArrayList<String> randInstructions = new ArrayList<String>(); ArrayList<String> recordReaderInstructions = new ArrayList<String>(); int numReducers = 0; int replication = 1; ArrayList<String> inputLabels = new ArrayList<String>(); ArrayList<String> outputLabels = new ArrayList<String>(); ArrayList<Instruction> renameInstructions = new ArrayList<Instruction>(); ArrayList<Instruction> variableInstructions = new ArrayList<Instruction>(); ArrayList<Instruction> postInstructions = new ArrayList<Instruction>(); ArrayList<Integer> MRJobLineNumbers = null; if (DMLScript.ENABLE_DEBUG_MODE) { MRJobLineNumbers = new ArrayList<Integer>(); } ArrayList<Lop> inputLops = new ArrayList<Lop>(); boolean cellModeOverride = false; /* Find the nodes that produce an output */ ArrayList<Lop> rootNodes = new ArrayList<Lop>(); getOutputNodes(execNodes, rootNodes, jt); if (LOG.isTraceEnabled()) LOG.trace("# of root nodes = " + rootNodes.size()); /* Remove transient writes that are simple copy of transient reads */ if (jt == JobType.GMR || jt == JobType.GMRCELL) { ArrayList<Lop> markedNodes = new ArrayList<Lop>(); // only keep data nodes that are results of some computation. for (Lop rnode : rootNodes) { if (rnode.getExecLocation() == ExecLocation.Data && ((Data) rnode).isTransient() && ((Data) rnode).getOperationType() == OperationTypes.WRITE && ((Data) rnode).getDataType() == DataType.MATRIX) { // no computation, just a copy if (rnode.getInputs().get(0).getExecLocation() == ExecLocation.Data && ((Data) rnode.getInputs().get(0)).isTransient() && rnode.getOutputParameters() .getLabel().equals(rnode.getInputs().get(0).getOutputParameters().getLabel())) { markedNodes.add(rnode); } } } // delete marked nodes rootNodes.removeAll(markedNodes); markedNodes.clear(); if (rootNodes.isEmpty()) return; } // structure that maps node to their indices that will be used in the instructions HashMap<Lop, Integer> nodeIndexMapping = new HashMap<Lop, Integer>(); /* Determine all input data files */ for (Lop rnode : rootNodes) { getInputPathsAndParameters(rnode, execNodes, inputs, inputInfos, numRows, numCols, numRowsPerBlock, numColsPerBlock, nodeIndexMapping, inputLabels, inputLops, MRJobLineNumbers); } // In case of RAND job, instructions are defined in the input file if (jt == JobType.DATAGEN) randInstructions = inputs; int[] start_index = new int[1]; start_index[0] = inputs.size(); /* Get RecordReader Instructions */ // currently, recordreader instructions are allowed only in GMR jobs if (jt == JobType.GMR || jt == JobType.GMRCELL) { for (Lop rnode : rootNodes) { getRecordReaderInstructions(rnode, execNodes, inputs, recordReaderInstructions, nodeIndexMapping, start_index, inputLabels, inputLops, MRJobLineNumbers); if (recordReaderInstructions.size() > 1) throw new LopsException("MapReduce job can only have a single recordreader instruction: " + recordReaderInstructions.toString()); } } /* * Handle cases when job's output is FORCED to be cell format. * - If there exist a cell input, then output can not be blocked. * Only exception is when jobType = REBLOCK/CSVREBLOCK (for obvisous reason) * or when jobType = RAND since RandJob takes a special input file, * whose format should not be used to dictate the output format. * - If there exists a recordReader instruction * - If jobtype = GroupedAgg. This job can only run in cell mode. */ // if (jt != JobType.REBLOCK && jt != JobType.CSV_REBLOCK && jt != JobType.DATAGEN && jt != JobType.TRANSFORM) { for (int i = 0; i < inputInfos.size(); i++) if (inputInfos.get(i) == InputInfo.BinaryCellInputInfo || inputInfos.get(i) == InputInfo.TextCellInputInfo) cellModeOverride = true; } if (!recordReaderInstructions.isEmpty() || jt == JobType.GROUPED_AGG) cellModeOverride = true; /* Get Mapper Instructions */ for (int i = 0; i < rootNodes.size(); i++) { getMapperInstructions(rootNodes.get(i), execNodes, inputs, mapperInstructions, nodeIndexMapping, start_index, inputLabels, inputLops, MRJobLineNumbers); } if (LOG.isTraceEnabled()) { LOG.trace(" Input strings: " + inputs.toString()); if (jt == JobType.DATAGEN) LOG.trace(" Rand instructions: " + getCSVString(randInstructions)); if (jt == JobType.GMR) LOG.trace(" RecordReader instructions: " + getCSVString(recordReaderInstructions)); LOG.trace(" Mapper instructions: " + getCSVString(mapperInstructions)); } /* Get Shuffle and Reducer Instructions */ ArrayList<String> shuffleInstructions = new ArrayList<String>(); ArrayList<String> aggInstructionsReducer = new ArrayList<String>(); ArrayList<String> otherInstructionsReducer = new ArrayList<String>(); for (Lop rn : rootNodes) { int resultIndex = getAggAndOtherInstructions(rn, execNodes, shuffleInstructions, aggInstructionsReducer, otherInstructionsReducer, nodeIndexMapping, start_index, inputLabels, inputLops, MRJobLineNumbers); if (resultIndex == -1) throw new LopsException("Unexpected error in piggybacking!"); if (rn.getExecLocation() == ExecLocation.Data && ((Data) rn).getOperationType() == Data.OperationTypes.WRITE && ((Data) rn).isTransient() && rootNodes.contains(rn.getInputs().get(0))) { // Both rn (a transient write) and its input are root nodes. // Instead of creating two copies of the data, simply generate a cpvar instruction NodeOutput out = setupNodeOutputs(rn, ExecType.MR, cellModeOverride, true); writeinst.addAll(out.getLastInstructions()); } else { resultIndices.add(Byte.valueOf((byte) resultIndex)); // setup output filenames and outputInfos and generate related instructions NodeOutput out = setupNodeOutputs(rn, ExecType.MR, cellModeOverride, false); outputLabels.add(out.getVarName()); outputs.add(out.getFileName()); outputInfos.add(out.getOutInfo()); if (LOG.isTraceEnabled()) { LOG.trace(" Output Info: " + out.getFileName() + ";" + OutputInfo.outputInfoToString(out.getOutInfo()) + ";" + out.getVarName()); } renameInstructions.addAll(out.getLastInstructions()); variableInstructions.addAll(out.getPreInstructions()); postInstructions.addAll(out.getPostInstructions()); } } /* Determine if the output dimensions are known */ byte[] resultIndicesByte = new byte[resultIndices.size()]; for (int i = 0; i < resultIndicesByte.length; i++) { resultIndicesByte[i] = resultIndices.get(i).byteValue(); } if (LOG.isTraceEnabled()) { LOG.trace(" Shuffle Instructions: " + getCSVString(shuffleInstructions)); LOG.trace(" Aggregate Instructions: " + getCSVString(aggInstructionsReducer)); LOG.trace(" Other instructions =" + getCSVString(otherInstructionsReducer)); LOG.trace(" Output strings: " + outputs.toString()); LOG.trace(" ResultIndices = " + resultIndices.toString()); } /* Prepare the MapReduce job instruction */ MRJobInstruction mr = new MRJobInstruction(jt); // check if this is a map-only job. If not, set the number of reducers if (!shuffleInstructions.isEmpty() || !aggInstructionsReducer.isEmpty() || !otherInstructionsReducer.isEmpty()) numReducers = total_reducers; // set inputs, outputs, and other other properties for the job mr.setInputOutputLabels(inputLabels.toArray(new String[0]), outputLabels.toArray(new String[0])); mr.setOutputs(resultIndicesByte); mr.setDimsUnknownFilePrefix(getFilePath()); mr.setNumberOfReducers(numReducers); mr.setReplication(replication); // set instructions for recordReader and mapper mr.setRecordReaderInstructions(getCSVString(recordReaderInstructions)); mr.setMapperInstructions(getCSVString(mapperInstructions)); //compute and set mapper memory requirements (for consistency of runtime piggybacking) if (jt == JobType.GMR) { double mem = 0; for (Lop n : execNodes) mem += computeFootprintInMapper(n); mr.setMemoryRequirements(mem); } if (jt == JobType.DATAGEN) mr.setRandInstructions(getCSVString(randInstructions)); // set shuffle instructions mr.setShuffleInstructions(getCSVString(shuffleInstructions)); // set reducer instruction mr.setAggregateInstructionsInReducer(getCSVString(aggInstructionsReducer)); mr.setOtherInstructionsInReducer(getCSVString(otherInstructionsReducer)); if (DMLScript.ENABLE_DEBUG_MODE) { // set line number information for each MR instruction mr.setMRJobInstructionsLineNumbers(MRJobLineNumbers); } /* Add the prepared instructions to output set */ inst.addAll(variableInstructions); inst.add(mr); inst.addAll(postInstructions); deleteinst.addAll(renameInstructions); for (Lop l : inputLops) { if (DMLScript.ENABLE_DEBUG_MODE) { processConsumers(l, rmvarinst, deleteinst, l); } else { processConsumers(l, rmvarinst, deleteinst, null); } } } /** * converts an array list into a Lop.INSTRUCTION_DELIMITOR separated string * * @param inputStrings list of input strings * @return Lop.INSTRUCTION_DELIMITOR separated string */ private static String getCSVString(ArrayList<String> inputStrings) { StringBuilder sb = new StringBuilder(); for (String str : inputStrings) { if (str != null) { if (sb.length() > 0) sb.append(Lop.INSTRUCTION_DELIMITOR); sb.append(str); } } return sb.toString(); } /** * Method to populate aggregate and other instructions in reducer. * * @param node low-level operator * @param execNodes list of exec nodes * @param shuffleInstructions list of shuffle instructions * @param aggInstructionsReducer ? * @param otherInstructionsReducer ? * @param nodeIndexMapping node index mapping * @param start_index start index * @param inputLabels list of input labels * @param inputLops list of input lops * @param MRJobLineNumbers MR job line numbers * @return -1 if problem * @throws LopsException if LopsException occurs */ private int getAggAndOtherInstructions(Lop node, ArrayList<Lop> execNodes, ArrayList<String> shuffleInstructions, ArrayList<String> aggInstructionsReducer, ArrayList<String> otherInstructionsReducer, HashMap<Lop, Integer> nodeIndexMapping, int[] start_index, ArrayList<String> inputLabels, ArrayList<Lop> inputLops, ArrayList<Integer> MRJobLineNumbers) throws LopsException { int ret_val = -1; if (nodeIndexMapping.containsKey(node)) return nodeIndexMapping.get(node); // if not an input source and not in exec nodes, return. if (!execNodes.contains(node)) return ret_val; ArrayList<Integer> inputIndices = new ArrayList<Integer>(); // recurse // For WRITE, since the first element from input is the real input (the other elements // are parameters for the WRITE operation), so we only need to take care of the // first element. if (node.getType() == Lop.Type.Data && ((Data) node).getOperationType() == Data.OperationTypes.WRITE) { ret_val = getAggAndOtherInstructions(node.getInputs().get(0), execNodes, shuffleInstructions, aggInstructionsReducer, otherInstructionsReducer, nodeIndexMapping, start_index, inputLabels, inputLops, MRJobLineNumbers); inputIndices.add(ret_val); } else { for (Lop cnode : node.getInputs()) { ret_val = getAggAndOtherInstructions(cnode, execNodes, shuffleInstructions, aggInstructionsReducer, otherInstructionsReducer, nodeIndexMapping, start_index, inputLabels, inputLops, MRJobLineNumbers); inputIndices.add(ret_val); } } if (node.getExecLocation() == ExecLocation.Data) { if (((Data) node).getFileFormatType() == FileFormatTypes.CSV && !(node.getInputs() .get(0) instanceof ParameterizedBuiltin && ((ParameterizedBuiltin) node.getInputs().get(0)) .getOp() == org.apache.sysml.lops.ParameterizedBuiltin.OperationTypes.TRANSFORM)) { // Generate write instruction, which goes into CSV_WRITE Job int output_index = start_index[0]; shuffleInstructions.add(node.getInstructions(inputIndices.get(0), output_index)); if (DMLScript.ENABLE_DEBUG_MODE) { MRJobLineNumbers.add(node._beginLine); } nodeIndexMapping.put(node, output_index); start_index[0]++; return output_index; } else { return ret_val; } } if (node.getExecLocation() == ExecLocation.MapAndReduce) { /* Generate Shuffle Instruction for "node", and return the index associated with produced output */ boolean instGenerated = true; int output_index = start_index[0]; switch (node.getType()) { /* Lop types that take a single input */ case ReBlock: case CSVReBlock: case SortKeys: case CentralMoment: case CoVariance: case GroupedAgg: case DataPartition: shuffleInstructions.add(node.getInstructions(inputIndices.get(0), output_index)); if (DMLScript.ENABLE_DEBUG_MODE) { MRJobLineNumbers.add(node._beginLine); } break; case ParameterizedBuiltin: if (((ParameterizedBuiltin) node) .getOp() == org.apache.sysml.lops.ParameterizedBuiltin.OperationTypes.TRANSFORM) { shuffleInstructions.add(node.getInstructions(output_index)); if (DMLScript.ENABLE_DEBUG_MODE) { MRJobLineNumbers.add(node._beginLine); } } break; /* Lop types that take two inputs */ case MMCJ: case MMRJ: case CombineBinary: shuffleInstructions .add(node.getInstructions(inputIndices.get(0), inputIndices.get(1), output_index)); if (DMLScript.ENABLE_DEBUG_MODE) { MRJobLineNumbers.add(node._beginLine); } break; /* Lop types that take three inputs */ case CombineTernary: shuffleInstructions.add(node.getInstructions(inputIndices.get(0), inputIndices.get(1), inputIndices.get(2), output_index)); if (DMLScript.ENABLE_DEBUG_MODE) { MRJobLineNumbers.add(node._beginLine); } break; default: instGenerated = false; break; } if (instGenerated) { nodeIndexMapping.put(node, output_index); start_index[0]++; return output_index; } else { return inputIndices.get(0); } } /* Get instructions for aligned reduce and other lops below the reduce. */ if (node.getExecLocation() == ExecLocation.Reduce || node.getExecLocation() == ExecLocation.MapOrReduce || hasChildNode(node, execNodes, ExecLocation.MapAndReduce)) { if (inputIndices.size() == 1) { int output_index = start_index[0]; start_index[0]++; if (node.getType() == Type.Aggregate) { aggInstructionsReducer.add(node.getInstructions(inputIndices.get(0), output_index)); if (DMLScript.ENABLE_DEBUG_MODE) { MRJobLineNumbers.add(node._beginLine); } } else { otherInstructionsReducer.add(node.getInstructions(inputIndices.get(0), output_index)); } if (DMLScript.ENABLE_DEBUG_MODE) { MRJobLineNumbers.add(node._beginLine); } nodeIndexMapping.put(node, output_index); return output_index; } else if (inputIndices.size() == 2) { int output_index = start_index[0]; start_index[0]++; otherInstructionsReducer .add(node.getInstructions(inputIndices.get(0), inputIndices.get(1), output_index)); if (DMLScript.ENABLE_DEBUG_MODE) { MRJobLineNumbers.add(node._beginLine); } nodeIndexMapping.put(node, output_index); // populate list of input labels. // only Unary lops can contribute to labels if (node instanceof Unary && node.getInputs().size() > 1) { int index = 0; for (int i = 0; i < node.getInputs().size(); i++) { if (node.getInputs().get(i).getDataType() == DataType.SCALAR) { index = i; break; } } if (node.getInputs().get(index).getExecLocation() == ExecLocation.Data && !((Data) (node.getInputs().get(index))).isLiteral()) { inputLabels.add(node.getInputs().get(index).getOutputParameters().getLabel()); inputLops.add(node.getInputs().get(index)); } if (node.getInputs().get(index).getExecLocation() != ExecLocation.Data) { inputLabels.add(node.getInputs().get(index).getOutputParameters().getLabel()); inputLops.add(node.getInputs().get(index)); } } return output_index; } else if (inputIndices.size() == 3 || node.getType() == Type.Ternary) { int output_index = start_index[0]; start_index[0]++; if (node.getType() == Type.Ternary) { // in case of CTABLE_TRANSFORM_SCALAR_WEIGHT: inputIndices.get(2) would be -1 otherInstructionsReducer.add(node.getInstructions(inputIndices.get(0), inputIndices.get(1), inputIndices.get(2), output_index)); if (DMLScript.ENABLE_DEBUG_MODE) { MRJobLineNumbers.add(node._beginLine); } nodeIndexMapping.put(node, output_index); } else if (node.getType() == Type.ParameterizedBuiltin) { otherInstructionsReducer.add(node.getInstructions(inputIndices.get(0), inputIndices.get(1), inputIndices.get(2), output_index)); if (DMLScript.ENABLE_DEBUG_MODE) { MRJobLineNumbers.add(node._beginLine); } nodeIndexMapping.put(node, output_index); } else { otherInstructionsReducer.add(node.getInstructions(inputIndices.get(0), inputIndices.get(1), inputIndices.get(2), output_index)); if (DMLScript.ENABLE_DEBUG_MODE) { MRJobLineNumbers.add(node._beginLine); } nodeIndexMapping.put(node, output_index); return output_index; } return output_index; } else if (inputIndices.size() == 4) { int output_index = start_index[0]; start_index[0]++; otherInstructionsReducer.add(node.getInstructions(inputIndices.get(0), inputIndices.get(1), inputIndices.get(2), inputIndices.get(3), output_index)); if (DMLScript.ENABLE_DEBUG_MODE) { MRJobLineNumbers.add(node._beginLine); } nodeIndexMapping.put(node, output_index); return output_index; } else throw new LopsException("Invalid number of inputs to a lop: " + inputIndices.size()); } return -1; } /** * Method to get record reader instructions for a MR job. * * @param node low-level operator * @param execNodes list of exec nodes * @param inputStrings list of input strings * @param recordReaderInstructions list of record reader instructions * @param nodeIndexMapping node index mapping * @param start_index start index * @param inputLabels list of input labels * @param inputLops list of input lops * @param MRJobLineNumbers MR job line numbers * @return -1 if problem * @throws LopsException if LopsException occurs */ private static int getRecordReaderInstructions(Lop node, ArrayList<Lop> execNodes, ArrayList<String> inputStrings, ArrayList<String> recordReaderInstructions, HashMap<Lop, Integer> nodeIndexMapping, int[] start_index, ArrayList<String> inputLabels, ArrayList<Lop> inputLops, ArrayList<Integer> MRJobLineNumbers) throws LopsException { // if input source, return index if (nodeIndexMapping.containsKey(node)) return nodeIndexMapping.get(node); // not input source and not in exec nodes, then return. if (!execNodes.contains(node)) return -1; ArrayList<Integer> inputIndices = new ArrayList<Integer>(); int max_input_index = -1; //N child_for_max_input_index = null; // get mapper instructions for (int i = 0; i < node.getInputs().size(); i++) { // recurse Lop childNode = node.getInputs().get(i); int ret_val = getRecordReaderInstructions(childNode, execNodes, inputStrings, recordReaderInstructions, nodeIndexMapping, start_index, inputLabels, inputLops, MRJobLineNumbers); inputIndices.add(ret_val); if (ret_val > max_input_index) { max_input_index = ret_val; //child_for_max_input_index = childNode; } } // only lops with execLocation as RecordReader can contribute // instructions if ((node.getExecLocation() == ExecLocation.RecordReader)) { int output_index = max_input_index; // cannot reuse index if this is true // need to add better indexing schemes output_index = start_index[0]; start_index[0]++; nodeIndexMapping.put(node, output_index); // populate list of input labels. // only Ranagepick lop can contribute to labels if (node.getType() == Type.PickValues) { PickByCount pbc = (PickByCount) node; if (pbc.getOperationType() == PickByCount.OperationTypes.RANGEPICK) { int scalarIndex = 1; // always the second input is a scalar // if data lop not a literal -- add label if (node.getInputs().get(scalarIndex).getExecLocation() == ExecLocation.Data && !((Data) (node.getInputs().get(scalarIndex))).isLiteral()) { inputLabels.add(node.getInputs().get(scalarIndex).getOutputParameters().getLabel()); inputLops.add(node.getInputs().get(scalarIndex)); } // if not data lop, then this is an intermediate variable. if (node.getInputs().get(scalarIndex).getExecLocation() != ExecLocation.Data) { inputLabels.add(node.getInputs().get(scalarIndex).getOutputParameters().getLabel()); inputLops.add(node.getInputs().get(scalarIndex)); } } } // get recordreader instruction. if (node.getInputs().size() == 2) { recordReaderInstructions .add(node.getInstructions(inputIndices.get(0), inputIndices.get(1), output_index)); if (DMLScript.ENABLE_DEBUG_MODE) { MRJobLineNumbers.add(node._beginLine); } } else throw new LopsException("Unexpected number of inputs while generating a RecordReader Instruction"); return output_index; } return -1; } /** * Method to get mapper instructions for a MR job. * * @param node low-level operator * @param execNodes list of exec nodes * @param inputStrings list of input strings * @param instructionsInMapper list of instructions in mapper * @param nodeIndexMapping ? * @param start_index starting index * @param inputLabels input labels * @param MRJoblineNumbers MR job line numbers * @return -1 if problem * @throws LopsException if LopsException occurs */ private int getMapperInstructions(Lop node, ArrayList<Lop> execNodes, ArrayList<String> inputStrings, ArrayList<String> instructionsInMapper, HashMap<Lop, Integer> nodeIndexMapping, int[] start_index, ArrayList<String> inputLabels, ArrayList<Lop> inputLops, ArrayList<Integer> MRJobLineNumbers) throws LopsException { // if input source, return index if (nodeIndexMapping.containsKey(node)) return nodeIndexMapping.get(node); // not input source and not in exec nodes, then return. if (!execNodes.contains(node)) return -1; ArrayList<Integer> inputIndices = new ArrayList<Integer>(); int max_input_index = -1; // get mapper instructions for (Lop childNode : node.getInputs()) { int ret_val = getMapperInstructions(childNode, execNodes, inputStrings, instructionsInMapper, nodeIndexMapping, start_index, inputLabels, inputLops, MRJobLineNumbers); inputIndices.add(ret_val); if (ret_val > max_input_index) { max_input_index = ret_val; } } // only map and map-or-reduce without a reduce child node can contribute // to mapper instructions. if ((node.getExecLocation() == ExecLocation.Map || node.getExecLocation() == ExecLocation.MapOrReduce) && !hasChildNode(node, execNodes, ExecLocation.MapAndReduce) && !hasChildNode(node, execNodes, ExecLocation.Reduce)) { int output_index = max_input_index; // cannot reuse index if this is true // need to add better indexing schemes // if (child_for_max_input_index.getOutputs().size() > 1) { output_index = start_index[0]; start_index[0]++; // } nodeIndexMapping.put(node, output_index); // populate list of input labels. // only Unary lops can contribute to labels if (node instanceof Unary && node.getInputs().size() > 1) { // Following code must be executed only for those Unary // operators that have more than one input // It should not be executed for "true" unary operators like // cos(A). int index = 0; for (int i1 = 0; i1 < node.getInputs().size(); i1++) { if (node.getInputs().get(i1).getDataType() == DataType.SCALAR) { index = i1; break; } } // if data lop not a literal -- add label if (node.getInputs().get(index).getExecLocation() == ExecLocation.Data && !((Data) (node.getInputs().get(index))).isLiteral()) { inputLabels.add(node.getInputs().get(index).getOutputParameters().getLabel()); inputLops.add(node.getInputs().get(index)); } // if not data lop, then this is an intermediate variable. if (node.getInputs().get(index).getExecLocation() != ExecLocation.Data) { inputLabels.add(node.getInputs().get(index).getOutputParameters().getLabel()); inputLops.add(node.getInputs().get(index)); } } // get mapper instruction. if (node.getInputs().size() == 1) instructionsInMapper.add(node.getInstructions(inputIndices.get(0), output_index)); else if (node.getInputs().size() == 2) { instructionsInMapper .add(node.getInstructions(inputIndices.get(0), inputIndices.get(1), output_index)); } else if (node.getInputs().size() == 3) instructionsInMapper.add(node.getInstructions(inputIndices.get(0), inputIndices.get(1), inputIndices.get(2), output_index)); else if (node.getInputs().size() == 4) { // Example: Reshape instructionsInMapper.add(node.getInstructions(inputIndices.get(0), inputIndices.get(1), inputIndices.get(2), inputIndices.get(3), output_index)); } else if (node.getInputs().size() == 5) { // Example: RangeBasedReIndex A[row_l:row_u, col_l:col_u] instructionsInMapper.add(node.getInstructions(inputIndices.get(0), inputIndices.get(1), inputIndices.get(2), inputIndices.get(3), inputIndices.get(4), output_index)); } else if (node.getInputs().size() == 7) { // Example: RangeBasedReIndex A[row_l:row_u, col_l:col_u] = B instructionsInMapper.add(node.getInstructions(inputIndices.get(0), inputIndices.get(1), inputIndices.get(2), inputIndices.get(3), inputIndices.get(4), inputIndices.get(5), inputIndices.get(6), output_index)); } else throw new LopsException( "Node with " + node.getInputs().size() + " inputs is not supported in dag.java."); if (DMLScript.ENABLE_DEBUG_MODE) { MRJobLineNumbers.add(node._beginLine); } return output_index; } return -1; } // Method to populate inputs and also populates node index mapping. private static void getInputPathsAndParameters(Lop node, ArrayList<Lop> execNodes, ArrayList<String> inputStrings, ArrayList<InputInfo> inputInfos, ArrayList<Long> numRows, ArrayList<Long> numCols, ArrayList<Long> numRowsPerBlock, ArrayList<Long> numColsPerBlock, HashMap<Lop, Integer> nodeIndexMapping, ArrayList<String> inputLabels, ArrayList<Lop> inputLops, ArrayList<Integer> MRJobLineNumbers) throws LopsException { // treat rand as an input. if (node.getType() == Type.DataGen && execNodes.contains(node) && !nodeIndexMapping.containsKey(node)) { numRows.add(node.getOutputParameters().getNumRows()); numCols.add(node.getOutputParameters().getNumCols()); numRowsPerBlock.add(node.getOutputParameters().getRowsInBlock()); numColsPerBlock.add(node.getOutputParameters().getColsInBlock()); inputStrings.add(node.getInstructions(inputStrings.size(), inputStrings.size())); if (DMLScript.ENABLE_DEBUG_MODE) { MRJobLineNumbers.add(node._beginLine); } inputInfos.add(InputInfo.TextCellInputInfo); nodeIndexMapping.put(node, inputStrings.size() - 1); return; } // get input file names if (!execNodes.contains(node) && !nodeIndexMapping.containsKey(node) && !(node.getExecLocation() == ExecLocation.Data) && (!(node.getExecLocation() == ExecLocation.ControlProgram && node.getDataType() == DataType.SCALAR)) || (!execNodes.contains(node) && node.getExecLocation() == ExecLocation.Data && ((Data) node).getOperationType() == Data.OperationTypes.READ && ((Data) node).getDataType() != DataType.SCALAR && !nodeIndexMapping.containsKey(node))) { if (node.getOutputParameters().getFile_name() != null) { inputStrings.add(node.getOutputParameters().getFile_name()); } else { // use label name inputStrings.add(Lop.VARIABLE_NAME_PLACEHOLDER + node.getOutputParameters().getLabel() + Lop.VARIABLE_NAME_PLACEHOLDER); } inputLabels.add(node.getOutputParameters().getLabel()); inputLops.add(node); numRows.add(node.getOutputParameters().getNumRows()); numCols.add(node.getOutputParameters().getNumCols()); numRowsPerBlock.add(node.getOutputParameters().getRowsInBlock()); numColsPerBlock.add(node.getOutputParameters().getColsInBlock()); InputInfo nodeInputInfo = null; // Check if file format type is binary or text and update infos if (node.getOutputParameters().isBlocked()) { if (node.getOutputParameters().getFormat() == Format.BINARY) nodeInputInfo = InputInfo.BinaryBlockInputInfo; else throw new LopsException("Invalid format (" + node.getOutputParameters().getFormat() + ") encountered for a node/lop (ID=" + node.getID() + ") with blocked output."); } else { if (node.getOutputParameters().getFormat() == Format.TEXT) nodeInputInfo = InputInfo.TextCellInputInfo; else nodeInputInfo = InputInfo.BinaryCellInputInfo; } /* * Hardcode output Key and Value Classes for SortKeys */ // TODO: statiko -- remove this hardcoding -- i.e., lops must encode // the information on key/value classes if (node.getType() == Type.SortKeys) { // SortKeys is the input to some other lop (say, L) // InputInfo of L is the ouputInfo of SortKeys, which is // (compactformat, doubleWriteable, IntWritable) nodeInputInfo = new InputInfo(PickFromCompactInputFormat.class, DoubleWritable.class, IntWritable.class); } else if (node.getType() == Type.CombineBinary) { // CombineBinary is the input to some other lop (say, L) // InputInfo of L is the ouputInfo of CombineBinary // And, the outputInfo of CombineBinary depends on the operation! CombineBinary combine = (CombineBinary) node; if (combine.getOperation() == org.apache.sysml.lops.CombineBinary.OperationTypes.PreSort) { nodeInputInfo = new InputInfo(SequenceFileInputFormat.class, DoubleWritable.class, IntWritable.class); } else if (combine .getOperation() == org.apache.sysml.lops.CombineBinary.OperationTypes.PreCentralMoment || combine .getOperation() == org.apache.sysml.lops.CombineBinary.OperationTypes.PreCovUnweighted || combine .getOperation() == org.apache.sysml.lops.CombineBinary.OperationTypes.PreGroupedAggUnweighted) { nodeInputInfo = InputInfo.WeightedPairInputInfo; } } else if (node.getType() == Type.CombineTernary) { nodeInputInfo = InputInfo.WeightedPairInputInfo; } inputInfos.add(nodeInputInfo); nodeIndexMapping.put(node, inputStrings.size() - 1); return; } // if exec nodes does not contain node at this point, return. if (!execNodes.contains(node)) return; // process children recursively for (Lop lop : node.getInputs()) { getInputPathsAndParameters(lop, execNodes, inputStrings, inputInfos, numRows, numCols, numRowsPerBlock, numColsPerBlock, nodeIndexMapping, inputLabels, inputLops, MRJobLineNumbers); } } /** * Method to find all terminal nodes. * * @param execNodes list of exec nodes * @param rootNodes list of root nodes * @param jt job type */ private static void getOutputNodes(ArrayList<Lop> execNodes, ArrayList<Lop> rootNodes, JobType jt) { for (Lop node : execNodes) { // terminal node if (node.getOutputs().isEmpty() && !rootNodes.contains(node)) { rootNodes.add(node); } else { // check for nodes with at least one child outside execnodes int cnt = 0; for (Lop lop : node.getOutputs()) { cnt += (!execNodes.contains(lop)) ? 1 : 0; } if (cnt > 0 && !rootNodes.contains(node) // not already a rootnode && !(node.getExecLocation() == ExecLocation.Data && ((Data) node).getOperationType() == OperationTypes.READ && ((Data) node).getDataType() == DataType.MATRIX)) // Not a matrix Data READ { if (jt.allowsSingleShuffleInstruction() && node.getExecLocation() != ExecLocation.MapAndReduce) continue; if (cnt < node.getOutputs().size()) { if (!node.getProducesIntermediateOutput()) rootNodes.add(node); } else rootNodes.add(node); } } } } /** * check to see if a is the child of b (i.e., there is a directed path from a to b) * * @param a child lop * @param b parent lop * @param IDMap id map * @return true if a child of b */ private static boolean isChild(Lop a, Lop b, HashMap<Long, Integer> IDMap) { int bID = IDMap.get(b.getID()); return a.get_reachable()[bID]; } /** * Method to topologically sort lops * * @param v list of lops */ @SuppressWarnings({ "unchecked", "rawtypes" }) private void doTopologicalSort_strict_order(ArrayList<Lop> v) { //int numNodes = v.size(); /* * Step 1: compute the level for each node in the DAG. Level for each node is * computed as lops are created. So, this step is need not be performed here. * Step 2: sort the nodes by level, and within a level by node ID. */ // Step1: Performed at the time of creating Lops // Step2: sort nodes by level, and then by node ID Lop[] nodearray = v.toArray(new Lop[0]); Arrays.sort(nodearray, new LopComparator()); // Copy sorted nodes into "v" and construct a mapping between Lop IDs and sequence of numbers v.clear(); IDMap.clear(); for (int i = 0; i < nodearray.length; i++) { v.add(nodearray[i]); IDMap.put(v.get(i).getID(), i); } /* * Compute of All-pair reachability graph (Transitive Closure) of the DAG. * - Perform a depth-first search (DFS) from every node $u$ in the DAG * - and construct the list of reachable nodes from the node $u$ * - store the constructed reachability information in $u$.reachable[] boolean array */ // // for (int i = 0; i < nodearray.length; i++) { boolean[] arr = v.get(i).create_reachable(nodearray.length); Arrays.fill(arr, false); dagDFS(v.get(i), arr); } // print the nodes in sorted order if (LOG.isTraceEnabled()) { for (Lop vnode : v) { StringBuilder sb = new StringBuilder(); sb.append(vnode.getID()); sb.append("("); sb.append(vnode.getLevel()); sb.append(") "); sb.append(vnode.getType()); sb.append("("); for (Lop vin : vnode.getInputs()) { sb.append(vin.getID()); sb.append(","); } sb.append("), "); LOG.trace(sb.toString()); } LOG.trace("topological sort -- done"); } } /** * Method to perform depth-first traversal from a given node in the DAG. * Store the reachability information in marked[] boolean array. * * @param root low-level operator * @param marked reachability results */ private void dagDFS(Lop root, boolean[] marked) { //contains check currently required for globalopt, will be removed when cleaned up if (!IDMap.containsKey(root.getID())) return; int mapID = IDMap.get(root.getID()); if (marked[mapID]) return; marked[mapID] = true; for (Lop lop : root.getOutputs()) { dagDFS(lop, marked); } } private static boolean hasDirectChildNode(Lop node, ArrayList<Lop> childNodes) { if (childNodes.isEmpty()) return false; for (Lop cnode : childNodes) { if (cnode.getOutputs().contains(node)) return true; } return false; } private boolean hasChildNode(Lop node, ArrayList<Lop> nodes) { return hasChildNode(node, nodes, ExecLocation.INVALID); } private boolean hasChildNode(Lop node, ArrayList<Lop> childNodes, ExecLocation type) { if (childNodes.isEmpty()) return false; int index = IDMap.get(node.getID()); for (Lop cnode : childNodes) { if ((type == ExecLocation.INVALID || cnode.getExecLocation() == type) && cnode.get_reachable()[index]) return true; } return false; } private Lop getChildNode(Lop node, ArrayList<Lop> childNodes, ExecLocation type) { if (childNodes.isEmpty()) return null; int index = IDMap.get(node.getID()); for (Lop cnode : childNodes) { if (cnode.getExecLocation() == type && cnode.get_reachable()[index]) return cnode; } return null; } /* * Returns a node "n" such that * 1) n \in parentNodes * 2) n is an ancestor of "node" * 3) n.ExecLocation = type * * Returns null if no such "n" exists * */ private Lop getParentNode(Lop node, ArrayList<Lop> parentNodes, ExecLocation type) { if (parentNodes.isEmpty()) return null; for (Lop pn : parentNodes) { int index = IDMap.get(pn.getID()); if (pn.getExecLocation() == type && node.get_reachable()[index]) return pn; } return null; } // Checks if "node" has any descendants in nodesVec with definedMRJob flag // set to true private boolean hasMRJobChildNode(Lop node, ArrayList<Lop> nodesVec) { if (nodesVec.isEmpty()) return false; int index = IDMap.get(node.getID()); for (Lop n : nodesVec) { if (n.definesMRJob() && n.get_reachable()[index]) return true; } return false; } private boolean checkDataGenAsChildNode(Lop node, ArrayList<Lop> nodesVec) { if (nodesVec.isEmpty()) return true; int index = IDMap.get(node.getID()); boolean onlyDatagen = true; for (Lop n : nodesVec) { if (n.definesMRJob() && n.get_reachable()[index] && JobType.findJobTypeFromLop(n) != JobType.DATAGEN) onlyDatagen = false; } // return true also when there is no lop in "nodesVec" that defines a MR job. return onlyDatagen; } private static int getChildAlignment(Lop node, ArrayList<Lop> execNodes, ExecLocation type) { for (Lop n : node.getInputs()) { if (!execNodes.contains(n)) continue; if (execNodes.contains(n) && n.getExecLocation() == type) { if (n.getBreaksAlignment()) return MR_CHILD_FOUND_BREAKS_ALIGNMENT; else return MR_CHILD_FOUND_DOES_NOT_BREAK_ALIGNMENT; } else { int ret = getChildAlignment(n, execNodes, type); if (ret == MR_CHILD_FOUND_DOES_NOT_BREAK_ALIGNMENT || ret == CHILD_DOES_NOT_BREAK_ALIGNMENT) { if (n.getBreaksAlignment()) return CHILD_BREAKS_ALIGNMENT; else return CHILD_DOES_NOT_BREAK_ALIGNMENT; } else if (ret == MRCHILD_NOT_FOUND || ret == CHILD_BREAKS_ALIGNMENT || ret == MR_CHILD_FOUND_BREAKS_ALIGNMENT) return ret; else throw new RuntimeException("Something wrong in getChildAlignment()."); } } return MRCHILD_NOT_FOUND; } private boolean hasParentNode(Lop node, ArrayList<Lop> parentNodes) { if (parentNodes.isEmpty()) return false; for (Lop pnode : parentNodes) { int index = IDMap.get(pnode.getID()); if (node.get_reachable()[index]) return true; } return false; } }