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

Example 11 with ParForStatementBlock

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

the class OptTreeConverter method rCreateAbstractOptNode.

public static OptNode rCreateAbstractOptNode(StatementBlock sb, ProgramBlock pb, LocalVariableMap vars, boolean topLevel, Set<String> memo) {
    OptNode node = null;
    if (pb instanceof IfProgramBlock && sb instanceof IfStatementBlock) {
        IfProgramBlock ipb = (IfProgramBlock) pb;
        IfStatementBlock isb = (IfStatementBlock) sb;
        IfStatement is = (IfStatement) isb.getStatement(0);
        node = new OptNode(NodeType.IF);
        _hlMap.putProgMapping(sb, pb, node);
        node.setExecType(ExecType.CP);
        node.setLineNumbers(isb.getBeginLine(), isb.getEndLine());
        // handle predicate
        isb.getPredicateHops().resetVisitStatus();
        node.addChilds(rCreateAbstractOptNodes(isb.getPredicateHops(), vars, memo));
        // process if branch
        OptNode ifn = new OptNode(NodeType.GENERIC);
        _hlMap.putProgMapping(sb, pb, ifn);
        ifn.setExecType(ExecType.CP);
        node.addChild(ifn);
        int len = is.getIfBody().size();
        for (int i = 0; i < ipb.getChildBlocksIfBody().size() && i < len; i++) {
            ProgramBlock lpb = ipb.getChildBlocksIfBody().get(i);
            StatementBlock lsb = is.getIfBody().get(i);
            ifn.addChild(rCreateAbstractOptNode(lsb, lpb, vars, false, memo));
        }
        // process else branch
        if (ipb.getChildBlocksElseBody() != null) {
            OptNode efn = new OptNode(NodeType.GENERIC);
            _hlMap.putProgMapping(sb, pb, efn);
            efn.setExecType(ExecType.CP);
            node.addChild(efn);
            int len2 = is.getElseBody().size();
            for (int i = 0; i < ipb.getChildBlocksElseBody().size() && i < len2; i++) {
                ProgramBlock lpb = ipb.getChildBlocksElseBody().get(i);
                StatementBlock lsb = is.getElseBody().get(i);
                efn.addChild(rCreateAbstractOptNode(lsb, lpb, vars, false, memo));
            }
        }
    } else if (pb instanceof WhileProgramBlock && sb instanceof WhileStatementBlock) {
        WhileProgramBlock wpb = (WhileProgramBlock) pb;
        WhileStatementBlock wsb = (WhileStatementBlock) sb;
        WhileStatement ws = (WhileStatement) wsb.getStatement(0);
        node = new OptNode(NodeType.WHILE);
        _hlMap.putProgMapping(sb, pb, node);
        node.setExecType(ExecType.CP);
        node.setLineNumbers(wsb.getBeginLine(), wsb.getEndLine());
        // handle predicate
        wsb.getPredicateHops().resetVisitStatus();
        node.addChilds(rCreateAbstractOptNodes(wsb.getPredicateHops(), vars, memo));
        // process body
        int len = ws.getBody().size();
        for (int i = 0; i < wpb.getChildBlocks().size() && i < len; i++) {
            ProgramBlock lpb = wpb.getChildBlocks().get(i);
            StatementBlock lsb = ws.getBody().get(i);
            node.addChild(rCreateAbstractOptNode(lsb, lpb, vars, false, memo));
        }
    } else if (pb instanceof ForProgramBlock && sb instanceof ForStatementBlock && !(pb instanceof ParForProgramBlock)) {
        ForProgramBlock fpb = (ForProgramBlock) pb;
        ForStatementBlock fsb = (ForStatementBlock) sb;
        ForStatement fs = (ForStatement) fsb.getStatement(0);
        node = new OptNode(NodeType.FOR);
        _hlMap.putProgMapping(sb, pb, node);
        node.setExecType(ExecType.CP);
        node.setLineNumbers(fsb.getBeginLine(), fsb.getEndLine());
        // determine number of iterations
        long N = OptimizerUtils.getNumIterations(fpb, vars, CostEstimator.FACTOR_NUM_ITERATIONS);
        node.addParam(ParamType.NUM_ITERATIONS, String.valueOf(N));
        // handle predicate
        fsb.getFromHops().resetVisitStatus();
        fsb.getToHops().resetVisitStatus();
        if (fsb.getIncrementHops() != null)
            fsb.getIncrementHops().resetVisitStatus();
        node.addChilds(rCreateAbstractOptNodes(fsb.getFromHops(), vars, memo));
        node.addChilds(rCreateAbstractOptNodes(fsb.getToHops(), vars, memo));
        if (fsb.getIncrementHops() != null)
            node.addChilds(rCreateAbstractOptNodes(fsb.getIncrementHops(), vars, memo));
        // process body
        int len = fs.getBody().size();
        for (int i = 0; i < fpb.getChildBlocks().size() && i < len; i++) {
            ProgramBlock lpb = fpb.getChildBlocks().get(i);
            StatementBlock lsb = fs.getBody().get(i);
            node.addChild(rCreateAbstractOptNode(lsb, lpb, vars, false, memo));
        }
    } else if (pb instanceof ParForProgramBlock && sb instanceof ParForStatementBlock) {
        ParForProgramBlock fpb = (ParForProgramBlock) pb;
        ParForStatementBlock fsb = (ParForStatementBlock) sb;
        ParForStatement fs = (ParForStatement) fsb.getStatement(0);
        node = new OptNode(NodeType.PARFOR);
        node.setLineNumbers(fsb.getBeginLine(), fsb.getEndLine());
        _hlMap.putProgMapping(sb, pb, node);
        node.setK(fpb.getDegreeOfParallelism());
        long N = fpb.getNumIterations();
        node.addParam(ParamType.NUM_ITERATIONS, (N != -1) ? String.valueOf(N) : String.valueOf(CostEstimator.FACTOR_NUM_ITERATIONS));
        switch(fpb.getExecMode()) {
            case LOCAL:
                node.setExecType(ExecType.CP);
                break;
            case REMOTE_MR:
            case REMOTE_MR_DP:
                node.setExecType(ExecType.MR);
                break;
            case REMOTE_SPARK:
            case REMOTE_SPARK_DP:
                node.setExecType(ExecType.SPARK);
                break;
            case UNSPECIFIED:
                node.setExecType(null);
        }
        if (!topLevel) {
            fsb.getFromHops().resetVisitStatus();
            fsb.getToHops().resetVisitStatus();
            if (fsb.getIncrementHops() != null)
                fsb.getIncrementHops().resetVisitStatus();
            node.addChilds(rCreateAbstractOptNodes(fsb.getFromHops(), vars, memo));
            node.addChilds(rCreateAbstractOptNodes(fsb.getToHops(), vars, memo));
            if (fsb.getIncrementHops() != null)
                node.addChilds(rCreateAbstractOptNodes(fsb.getIncrementHops(), vars, memo));
        }
        // process body
        int len = fs.getBody().size();
        for (int i = 0; i < fpb.getChildBlocks().size() && i < len; i++) {
            ProgramBlock lpb = fpb.getChildBlocks().get(i);
            StatementBlock lsb = fs.getBody().get(i);
            node.addChild(rCreateAbstractOptNode(lsb, lpb, vars, false, memo));
        }
        // parameters, add required parameters
        Map<String, String> lparams = fpb.getParForParams();
        node.addParam(ParamType.DATA_PARTITIONER, lparams.get(ParForStatementBlock.DATA_PARTITIONER));
        node.addParam(ParamType.TASK_PARTITIONER, lparams.get(ParForStatementBlock.TASK_PARTITIONER));
        node.addParam(ParamType.RESULT_MERGE, lparams.get(ParForStatementBlock.RESULT_MERGE));
    // TODO task size
    } else // last level program block
    {
        sb = pb.getStatementBlock();
        // process all hops
        node = new OptNode(NodeType.GENERIC);
        _hlMap.putProgMapping(sb, pb, node);
        node.addChilds(createAbstractOptNodes(sb.getHops(), vars, memo));
        node.setExecType(ExecType.CP);
        node.setLineNumbers(sb.getBeginLine(), sb.getEndLine());
        // TODO remove this workaround once this information can be obtained from hops/lops compiler
        if (node.isCPOnly()) {
            boolean isSparkExec = OptimizerUtils.isSparkExecutionMode();
            if (!isSparkExec && containsMRJobInstruction(pb, false, false))
                node.setExecType(ExecType.MR);
            else if (isSparkExec && containsMRJobInstruction(pb, false, true))
                node.setExecType(ExecType.SPARK);
        }
    }
    // final cleanup
    // NOTE: required because this function is also used to create subtrees
    node.checkAndCleanupLeafNodes();
    return node;
}

Example 12 with ParForStatementBlock

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

the class OptimizerRuleBased method rewriteInjectSparkLoopCheckpointing.

// /////
// REWRITE inject spark loop checkpointing
// /
protected void rewriteInjectSparkLoopCheckpointing(OptNode n) {
    // get program blocks of root parfor
    Object[] progobj = OptTreeConverter.getAbstractPlanMapping().getMappedProg(n.getID());
    ParForStatementBlock pfsb = (ParForStatementBlock) progobj[0];
    ParForStatement fs = (ParForStatement) pfsb.getStatement(0);
    ParForProgramBlock pfpb = (ParForProgramBlock) progobj[1];
    boolean applied = false;
    try {
        // apply hop rewrite inject spark checkpoints (but without context awareness)
        RewriteInjectSparkLoopCheckpointing rewrite = new RewriteInjectSparkLoopCheckpointing(false);
        ProgramRewriter rewriter = new ProgramRewriter(rewrite);
        ProgramRewriteStatus state = new ProgramRewriteStatus();
        rewriter.rRewriteStatementBlockHopDAGs(pfsb, state);
        fs.setBody(rewriter.rRewriteStatementBlocks(fs.getBody(), state, true));
        // recompile if additional checkpoints introduced
        if (state.getInjectedCheckpoints()) {
            pfpb.setChildBlocks(ProgramRecompiler.generatePartitialRuntimeProgram(pfpb.getProgram(), fs.getBody()));
            applied = true;
        }
    } catch (Exception ex) {
        throw new DMLRuntimeException(ex);
    }
    LOG.debug(getOptMode() + " OPT: rewrite 'inject spark loop checkpointing' - result=" + applied);
}

Example 13 with ParForStatementBlock

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

the class OptimizerRuleBased method rewriteInjectSparkRepartition.

// /////
// REWRITE inject spark repartition for zipmm
// /
protected void rewriteInjectSparkRepartition(OptNode n, LocalVariableMap vars) {
    // get program blocks of root parfor
    Object[] progobj = OptTreeConverter.getAbstractPlanMapping().getMappedProg(n.getID());
    ParForStatementBlock pfsb = (ParForStatementBlock) progobj[0];
    ParForProgramBlock pfpb = (ParForProgramBlock) progobj[1];
    ArrayList<String> ret = new ArrayList<>();
    if (// spark exec mode
    OptimizerUtils.isSparkExecutionMode() && // local parfor
    n.getExecType() == ExecType.CP && // at least 2 iterations
    _N > 1) {
        // collect candidates from zipmm spark instructions
        HashSet<String> cand = new HashSet<>();
        rCollectZipmmPartitioningCandidates(n, cand);
        // prune updated candidates
        HashSet<String> probe = new HashSet<>(pfsb.getReadOnlyParentVars());
        for (String var : cand) if (probe.contains(var))
            ret.add(var);
        // prune small candidates
        ArrayList<String> tmp = new ArrayList<>(ret);
        ret.clear();
        for (String var : tmp) if (vars.get(var) instanceof MatrixObject) {
            MatrixObject mo = (MatrixObject) vars.get(var);
            double sp = OptimizerUtils.getSparsity(mo.getNumRows(), mo.getNumColumns(), mo.getNnz());
            double size = OptimizerUtils.estimateSizeExactSparsity(mo.getNumRows(), mo.getNumColumns(), sp);
            if (size > OptimizerUtils.getLocalMemBudget())
                ret.add(var);
        }
        // apply rewrite to parfor pb
        if (!ret.isEmpty()) {
            pfpb.setSparkRepartitionVariables(ret);
        }
    }
    _numEvaluatedPlans++;
    LOG.debug(getOptMode() + " OPT: rewrite 'inject spark input repartition' - result=" + ret.size() + " (" + ProgramConverter.serializeStringCollection(ret) + ")");
}

Example 14 with ParForStatementBlock

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

the class OptimizerRuleBased method rewriteSetDataPartitioner.

// /////
// REWRITE set data partitioner
// /
protected boolean rewriteSetDataPartitioner(OptNode n, LocalVariableMap vars, HashMap<String, PartitionFormat> partitionedMatrices, double thetaM) {
    if (n.getNodeType() != NodeType.PARFOR)
        LOG.warn(getOptMode() + " OPT: Data partitioner can only be set for a ParFor node.");
    boolean blockwise = false;
    // preparations
    long id = n.getID();
    Object[] o = OptTreeConverter.getAbstractPlanMapping().getMappedProg(id);
    ParForStatementBlock pfsb = (ParForStatementBlock) o[0];
    ParForProgramBlock pfpb = (ParForProgramBlock) o[1];
    // search for candidates
    boolean apply = false;
    if (// only if we are allowed to recompile
    OptimizerUtils.isHybridExecutionMode() && // only if beneficial wrt problem size
    (_N >= PROB_SIZE_THRESHOLD_PARTITIONING || _Nmax >= PROB_SIZE_THRESHOLD_PARTITIONING)) {
        HashMap<String, PartitionFormat> cand2 = new HashMap<>();
        for (String c : pfsb.getReadOnlyParentVars()) {
            PartitionFormat dpf = pfsb.determineDataPartitionFormat(c);
            if (dpf != PartitionFormat.NONE && dpf._dpf != PDataPartitionFormat.BLOCK_WISE_M_N) {
                cand2.put(c, dpf);
            }
        }
        apply = rFindDataPartitioningCandidates(n, cand2, vars, thetaM);
        if (apply)
            partitionedMatrices.putAll(cand2);
    }
    PDataPartitioner REMOTE = OptimizerUtils.isSparkExecutionMode() ? PDataPartitioner.REMOTE_SPARK : PDataPartitioner.REMOTE_MR;
    PDataPartitioner pdp = (apply) ? REMOTE : PDataPartitioner.NONE;
    // NOTE: since partitioning is only applied in case of MR index access, we assume a large
    // matrix and hence always apply REMOTE_MR (the benefit for large matrices outweigths
    // potentially unnecessary MR jobs for smaller matrices)
    // modify rtprog
    pfpb.setDataPartitioner(pdp);
    // modify plan
    n.addParam(ParamType.DATA_PARTITIONER, pdp.toString());
    _numEvaluatedPlans++;
    LOG.debug(getOptMode() + " OPT: rewrite 'set data partitioner' - result=" + pdp.toString() + " (" + ProgramConverter.serializeStringCollection(partitionedMatrices.keySet()) + ")");
    return blockwise;
}

Example 15 with ParForStatementBlock

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

the class ParForProgramBlock method executeRemoteMRParForDP.

private void executeRemoteMRParForDP(ExecutionContext ec, IntObject itervar, IntObject from, IntObject to, IntObject incr) throws IOException {
    /* Step 0) check and recompile MR inst
		 * Step 1) serialize child PB and inst
		 * Step 2) create and serialize tasks
		 * Step 3) submit MR Jobs and wait for results
		 * Step 4) collect results from each parallel worker
		 */
    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);
    String program = ProgramConverter.serializeParForBody(body);
    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
    exportMatricesToHDFS(ec);
    // Step 4) submit MR job (wait for finished work)
    OutputInfo inputOI = ((inputMatrix.getSparsity() < 0.1 && inputDPF == PartitionFormat.COLUMN_WISE) || (inputMatrix.getSparsity() < 0.001 && inputDPF == PartitionFormat.ROW_WISE)) ? OutputInfo.BinaryCellOutputInfo : OutputInfo.BinaryBlockOutputInfo;
    RemoteParForJobReturn ret = RemoteDPParForMR.runJob(_ID, _iterPredVar, _colocatedDPMatrix, program, resultFile, inputMatrix, inputDPF, inputOI, _tSparseCol, _enableCPCaching, _numThreads, _replicationDP);
    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);
    }
}