Examples with ParForStatementBlock org.apache.sysml.parser.ParForStatementBlock used on opensource projects

Example 16 with ParForStatementBlock

use of org.apache.sysml.parser.ParForStatementBlock in project incubator-systemml by apache.

the class ParForProgramBlock method getMinMemory.

private long getMinMemory(ExecutionContext ec) {
    long ret = -1;
    // if forced remote exec and single node
    if (DMLScript.rtplatform == RUNTIME_PLATFORM.SINGLE_NODE && _execMode == PExecMode.REMOTE_MR && _optMode == POptMode.NONE) {
        try {
            ParForStatementBlock sb = (ParForStatementBlock) getStatementBlock();
            OptTree tree = OptTreeConverter.createAbstractOptTree(-1, -1, sb, this, new HashSet<String>(), ec);
            CostEstimator est = new CostEstimatorHops(OptTreeConverter.getAbstractPlanMapping());
            double mem = est.getEstimate(TestMeasure.MEMORY_USAGE, tree.getRoot());
            ret = (long) (mem * (1d / OptimizerUtils.MEM_UTIL_FACTOR));
        } catch (Exception e) {
            LOG.error("Failed to analyze minmum memory requirements.", e);
        }
    }
    return ret;
}

Example 17 with ParForStatementBlock

use of org.apache.sysml.parser.ParForStatementBlock in project incubator-systemml by apache.

the class ParForProgramBlock method execute.

@Override
public void execute(ExecutionContext ec) {
    ParForStatementBlock sb = (ParForStatementBlock) getStatementBlock();
    // evaluate from, to, incr only once (assumption: known at for entry)
    IntObject from = executePredicateInstructions(1, _fromInstructions, ec);
    IntObject to = executePredicateInstructions(2, _toInstructions, ec);
    IntObject incr = (_incrementInstructions == null || _incrementInstructions.isEmpty()) ? new IntObject((from.getLongValue() <= to.getLongValue()) ? 1 : -1) : executePredicateInstructions(3, _incrementInstructions, ec);
    if (// would produce infinite loop
    incr.getLongValue() == 0)
        throw new DMLRuntimeException(this.printBlockErrorLocation() + "Expression for increment " + "of variable '" + _iterPredVar + "' must evaluate to a non-zero value.");
    // early exit on num iterations = zero
    _numIterations = computeNumIterations(from, to, incr);
    if (_numIterations <= 0)
        // avoid unnecessary optimization/initialization
        return;
    // /////
    if (_optMode != POptMode.NONE) {
        // set optimizer log level
        OptimizationWrapper.setLogLevel(_optLogLevel);
        // core optimize
        OptimizationWrapper.optimize(_optMode, sb, this, ec, _monitor);
    }
    // /////
    // DATA PARTITIONING of read-only parent variables of type (matrix,unpartitioned)
    // /////
    Timing time = _monitor ? new Timing(true) : null;
    // partitioning on demand (note: for fused data partitioning and execute the optimizer set
    // the data partitioner to NONE in order to prevent any side effects)
    handleDataPartitioning(ec);
    // repartitioning of variables for spark cpmm/zipmm in order prevent unnecessary shuffle
    handleSparkRepartitioning(ec);
    // eager rdd caching of variables for spark in order prevent read/write contention
    handleSparkEagerCaching(ec);
    if (_monitor)
        StatisticMonitor.putPFStat(_ID, Stat.PARFOR_INIT_DATA_T, time.stop());
    // initialize iter var to form value
    IntObject iterVar = new IntObject(from.getLongValue());
    // /////
    // begin PARALLEL EXECUTION of (PAR)FOR body
    // /////
    LOG.trace("EXECUTE PARFOR ID = " + _ID + " with mode = " + _execMode + ", numThreads = " + _numThreads + ", taskpartitioner = " + _taskPartitioner);
    if (_monitor) {
        StatisticMonitor.putPFStat(_ID, Stat.PARFOR_NUMTHREADS, _numThreads);
        StatisticMonitor.putPFStat(_ID, Stat.PARFOR_TASKSIZE, _taskSize);
        StatisticMonitor.putPFStat(_ID, Stat.PARFOR_TASKPARTITIONER, _taskPartitioner.ordinal());
        StatisticMonitor.putPFStat(_ID, Stat.PARFOR_DATAPARTITIONER, _dataPartitioner.ordinal());
        StatisticMonitor.putPFStat(_ID, Stat.PARFOR_EXECMODE, _execMode.ordinal());
    }
    // preserve shared input/result variables of cleanup
    ArrayList<String> varList = ec.getVarList();
    boolean[] varState = ec.pinVariables(varList);
    try {
        switch(_execMode) {
            case // create parworkers as local threads
            LOCAL:
                executeLocalParFor(ec, iterVar, from, to, incr);
                break;
            case // create parworkers as MR tasks (one job per parfor)
            REMOTE_MR:
                executeRemoteMRParFor(ec, iterVar, from, to, incr);
                break;
            case // create parworkers as MR tasks (one job per parfor)
            REMOTE_MR_DP:
                executeRemoteMRParForDP(ec, iterVar, from, to, incr);
                break;
            case // create parworkers as Spark tasks (one job per parfor)
            REMOTE_SPARK:
                executeRemoteSparkParFor(ec, iterVar, from, to, incr);
                break;
            case // create parworkers as Spark tasks (one job per parfor)
            REMOTE_SPARK_DP:
                executeRemoteSparkParForDP(ec, iterVar, from, to, incr);
                break;
            default:
                throw new DMLRuntimeException("Undefined execution mode: '" + _execMode + "'.");
        }
    } catch (Exception ex) {
        throw new DMLRuntimeException("PARFOR: Failed to execute loop in parallel.", ex);
    }
    // reset state of shared input/result variables
    ec.unpinVariables(varList, varState);
    // cleanup unpinned shared variables
    cleanupSharedVariables(ec, varState);
    // set iteration var to TO value (+ increment) for FOR equivalence
    // consistent with for
    iterVar = new IntObject(to.getLongValue());
    ec.setVariable(_iterPredVar, iterVar);
    // we can replace those variables, because partitioning only applied for read-only matrices
    for (String var : _variablesDPOriginal.keySet()) {
        // cleanup partitioned matrix (if not reused)
        if (!_variablesDPReuse.keySet().contains(var))
            VariableCPInstruction.processRemoveVariableInstruction(ec, var);
        // reset to original matrix
        MatrixObject mo = (MatrixObject) _variablesDPOriginal.get(var);
        ec.setVariable(var, mo);
    }
    // print profiling report (only if top-level parfor because otherwise in parallel context)
    if (_monitorReport)
        LOG.info("\n" + StatisticMonitor.createReport());
    // TODO reset of hop parallelism constraint (e.g., ba+*)
    for (// release forced exectypes
    String dpvar : // release forced exectypes
    _variablesDPOriginal.keySet()) ProgramRecompiler.rFindAndRecompileIndexingHOP(sb, this, dpvar, ec, false);
    // release forced exectypes for fused dp/exec
    if (_execMode == PExecMode.REMOTE_MR_DP || _execMode == PExecMode.REMOTE_SPARK_DP)
        ProgramRecompiler.rFindAndRecompileIndexingHOP(sb, this, _colocatedDPMatrix, ec, false);
    // after release, deletes dp_varnames
    resetOptimizerFlags();
    // execute exit instructions (usually empty)
    executeInstructions(_exitInstructions, ec);
}

Example 18 with ParForStatementBlock

use of org.apache.sysml.parser.ParForStatementBlock in project incubator-systemml by apache.

the class ParForProgramBlock method executeRemoteSparkParForDP.

private void executeRemoteSparkParForDP(ExecutionContext ec, IntObject itervar, IntObject from, IntObject to, IntObject incr) throws IOException {
    Timing time = (_monitor ? new Timing(true) : null);
    // Step 0) check and compile to CP (if forced remote parfor)
    boolean flagForced = checkMRAndRecompileToCP(0);
    // Step 1) prepare partitioned input matrix (needs to happen before serializing the program)
    ParForStatementBlock sb = (ParForStatementBlock) getStatementBlock();
    MatrixObject inputMatrix = ec.getMatrixObject(_colocatedDPMatrix);
    PartitionFormat inputDPF = sb.determineDataPartitionFormat(_colocatedDPMatrix);
    // mark matrix var as partitioned
    inputMatrix.setPartitioned(inputDPF._dpf, inputDPF._N);
    // Step 2) init parallel workers (serialize PBs)
    // NOTES: each mapper changes filenames with regard to his ID as we submit a single
    // job, cannot reuse serialized string, since variables are serialized as well.
    ParForBody body = new ParForBody(_childBlocks, _resultVars, ec);
    HashMap<String, byte[]> clsMap = new HashMap<>();
    String program = ProgramConverter.serializeParForBody(body, clsMap);
    if (_monitor)
        StatisticMonitor.putPFStat(_ID, Stat.PARFOR_INIT_PARWRK_T, time.stop());
    // Step 3) create tasks
    TaskPartitioner partitioner = createTaskPartitioner(from, to, incr);
    String resultFile = constructResultFileName();
    long numIterations = partitioner.getNumIterations();
    // partitioner.createTasks().size();
    long numCreatedTasks = numIterations;
    if (_monitor)
        StatisticMonitor.putPFStat(_ID, Stat.PARFOR_INIT_TASKS_T, time.stop());
    // write matrices to HDFS, except DP matrix which is the input to the RemoteDPParForSpark job
    exportMatricesToHDFS(ec, _colocatedDPMatrix);
    // Step 4) submit MR job (wait for finished work)
    // TODO runtime support for binary cell partitioning
    OutputInfo inputOI = OutputInfo.BinaryBlockOutputInfo;
    RemoteParForJobReturn ret = RemoteDPParForSpark.runJob(_ID, _iterPredVar, _colocatedDPMatrix, program, clsMap, resultFile, inputMatrix, ec, inputDPF, inputOI, _tSparseCol, _enableCPCaching, _numThreads);
    if (_monitor)
        StatisticMonitor.putPFStat(_ID, Stat.PARFOR_WAIT_EXEC_T, time.stop());
    // Step 5) collecting results from each parallel worker
    int numExecutedTasks = ret.getNumExecutedTasks();
    int numExecutedIterations = ret.getNumExecutedIterations();
    // consolidate results into global symbol table
    consolidateAndCheckResults(ec, numIterations, numCreatedTasks, numExecutedIterations, numExecutedTasks, ret.getVariables());
    if (// see step 0
    flagForced)
        releaseForcedRecompile(0);
    inputMatrix.unsetPartitioned();
    if (_monitor) {
        StatisticMonitor.putPFStat(_ID, Stat.PARFOR_WAIT_RESULTS_T, time.stop());
        StatisticMonitor.putPFStat(_ID, Stat.PARFOR_NUMTASKS, numExecutedTasks);
        StatisticMonitor.putPFStat(_ID, Stat.PARFOR_NUMITERS, numExecutedIterations);
    }
}

Example 19 with ParForStatementBlock

use of org.apache.sysml.parser.ParForStatementBlock in project incubator-systemml by apache.

the class ParForProgramBlock method handleDataPartitioning.

private void handleDataPartitioning(ExecutionContext ec) {
    PDataPartitioner dataPartitioner = _dataPartitioner;
    if (dataPartitioner != PDataPartitioner.NONE) {
        ParForStatementBlock sb = (ParForStatementBlock) getStatementBlock();
        if (sb == null)
            throw new DMLRuntimeException("ParFor statement block required for reasoning about data partitioning.");
        for (String var : sb.getReadOnlyParentVars()) {
            Data dat = ec.getVariable(var);
            // partitioning but typically related branches are never executed)
            if (dat != null && dat instanceof MatrixObject) {
                // unpartitioned input
                MatrixObject moVar = (MatrixObject) dat;
                PartitionFormat dpf = sb.determineDataPartitionFormat(var);
                LOG.trace("PARFOR ID = " + _ID + ", Partitioning read-only input variable " + var + " (format=" + dpf + ", mode=" + _dataPartitioner + ")");
                if (dpf != PartitionFormat.NONE) {
                    if (dataPartitioner != PDataPartitioner.REMOTE_SPARK && dpf.isBlockwise()) {
                        LOG.warn("PARFOR ID = " + _ID + ", Switching data partitioner from " + dataPartitioner + " to " + PDataPartitioner.REMOTE_SPARK.name() + " for blockwise-n partitioning.");
                        dataPartitioner = PDataPartitioner.REMOTE_SPARK;
                    }
                    Timing ltime = new Timing(true);
                    // input data partitioning (reuse if possible)
                    Data dpdatNew = _variablesDPReuse.get(var);
                    if (// no reuse opportunity
                    dpdatNew == null) {
                        DataPartitioner dp = createDataPartitioner(dpf, dataPartitioner, ec);
                        // disable binary cell for sparse if consumed by MR jobs
                        if (!OptimizerRuleBased.allowsBinaryCellPartitions(moVar, dpf) || // TODO support for binarycell
                        OptimizerUtils.isSparkExecutionMode()) {
                            dp.disableBinaryCell();
                        }
                        MatrixObject moVarNew = dp.createPartitionedMatrixObject(moVar, constructDataPartitionsFileName());
                        dpdatNew = moVarNew;
                        // skip remaining partitioning logic if not partitioned (e.g., too small)
                        if (moVar == moVarNew)
                            // skip to next
                            continue;
                    }
                    ec.setVariable(var, dpdatNew);
                    // recompile parfor body program
                    ProgramRecompiler.rFindAndRecompileIndexingHOP(sb, this, var, ec, true);
                    // store original and partitioned matrix (for reuse if applicable)
                    _variablesDPOriginal.put(var, moVar);
                    if (ALLOW_REUSE_PARTITION_VARS && ProgramRecompiler.isApplicableForReuseVariable(sb.getDMLProg(), sb, var)) {
                        _variablesDPReuse.put(var, dpdatNew);
                    }
                    LOG.trace("Partitioning and recompilation done in " + ltime.stop() + "ms");
                }
            }
        }
    }
}

Example 20 with ParForStatementBlock

use of org.apache.sysml.parser.ParForStatementBlock in project incubator-systemml by apache.

the class Explain method getHopDAG.

private static StringBuilder getHopDAG(StatementBlock sb, StringBuilder nodes, ArrayList<Integer> lines, boolean withSubgraph) {
    StringBuilder builder = new StringBuilder();
    if (sb instanceof WhileStatementBlock) {
        addSubGraphHeader(builder, withSubgraph);
        WhileStatementBlock wsb = (WhileStatementBlock) sb;
        String label = null;
        if (!wsb.getUpdateInPlaceVars().isEmpty())
            label = "WHILE (lines " + wsb.getBeginLine() + "-" + wsb.getEndLine() + ") in-place=" + wsb.getUpdateInPlaceVars().toString() + "";
        else
            label = "WHILE (lines " + wsb.getBeginLine() + "-" + wsb.getEndLine() + ")";
        // TODO: Don't show predicate hops for now
        // builder.append(explainHop(wsb.getPredicateHops()));
        WhileStatement ws = (WhileStatement) sb.getStatement(0);
        for (StatementBlock current : ws.getBody()) builder.append(getHopDAG(current, nodes, lines, withSubgraph));
        addSubGraphFooter(builder, withSubgraph, label);
    } else if (sb instanceof IfStatementBlock) {
        addSubGraphHeader(builder, withSubgraph);
        IfStatementBlock ifsb = (IfStatementBlock) sb;
        String label = "IF (lines " + ifsb.getBeginLine() + "-" + ifsb.getEndLine() + ")";
        // TODO: Don't show predicate hops for now
        // builder.append(explainHop(ifsb.getPredicateHops(), level+1));
        IfStatement ifs = (IfStatement) sb.getStatement(0);
        for (StatementBlock current : ifs.getIfBody()) {
            builder.append(getHopDAG(current, nodes, lines, withSubgraph));
            addSubGraphFooter(builder, withSubgraph, label);
        }
        if (!ifs.getElseBody().isEmpty()) {
            addSubGraphHeader(builder, withSubgraph);
            label = "ELSE (lines " + ifsb.getBeginLine() + "-" + ifsb.getEndLine() + ")";
            for (StatementBlock current : ifs.getElseBody()) builder.append(getHopDAG(current, nodes, lines, withSubgraph));
            addSubGraphFooter(builder, withSubgraph, label);
        }
    } else if (sb instanceof ForStatementBlock) {
        ForStatementBlock fsb = (ForStatementBlock) sb;
        addSubGraphHeader(builder, withSubgraph);
        String label = "";
        if (sb instanceof ParForStatementBlock) {
            if (!fsb.getUpdateInPlaceVars().isEmpty())
                label = "PARFOR (lines " + fsb.getBeginLine() + "-" + fsb.getEndLine() + ") in-place=" + fsb.getUpdateInPlaceVars().toString() + "";
            else
                label = "PARFOR (lines " + fsb.getBeginLine() + "-" + fsb.getEndLine() + ")";
        } else {
            if (!fsb.getUpdateInPlaceVars().isEmpty())
                label = "FOR (lines " + fsb.getBeginLine() + "-" + fsb.getEndLine() + ") in-place=" + fsb.getUpdateInPlaceVars().toString() + "";
            else
                label = "FOR (lines " + fsb.getBeginLine() + "-" + fsb.getEndLine() + ")";
        }
        // TODO: Don't show predicate hops for now
        // if (fsb.getFromHops() != null)
        // builder.append(explainHop(fsb.getFromHops(), level+1));
        // if (fsb.getToHops() != null)
        // builder.append(explainHop(fsb.getToHops(), level+1));
        // if (fsb.getIncrementHops() != null)
        // builder.append(explainHop(fsb.getIncrementHops(), level+1));
        ForStatement fs = (ForStatement) sb.getStatement(0);
        for (StatementBlock current : fs.getBody()) builder.append(getHopDAG(current, nodes, lines, withSubgraph));
        addSubGraphFooter(builder, withSubgraph, label);
    } else if (sb instanceof FunctionStatementBlock) {
        FunctionStatement fsb = (FunctionStatement) sb.getStatement(0);
        addSubGraphHeader(builder, withSubgraph);
        String label = "Function (lines " + fsb.getBeginLine() + "-" + fsb.getEndLine() + ")";
        for (StatementBlock current : fsb.getBody()) builder.append(getHopDAG(current, nodes, lines, withSubgraph));
        addSubGraphFooter(builder, withSubgraph, label);
    } else {
        // For generic StatementBlock
        if (sb.requiresRecompilation()) {
            addSubGraphHeader(builder, withSubgraph);
        }
        ArrayList<Hop> hopsDAG = sb.getHops();
        if (hopsDAG != null && !hopsDAG.isEmpty()) {
            Hop.resetVisitStatus(hopsDAG);
            for (Hop hop : hopsDAG) builder.append(getHopDAG(hop, nodes, lines, withSubgraph));
            Hop.resetVisitStatus(hopsDAG);
        }
        if (sb.requiresRecompilation()) {
            builder.append("style=filled;\n");
            builder.append("color=lightgrey;\n");
            String label = "(lines " + sb.getBeginLine() + "-" + sb.getEndLine() + ") [recompile=" + sb.requiresRecompilation() + "]";
            addSubGraphFooter(builder, withSubgraph, label);
        }
    }
    return builder;
}