Examples with PExecMode org.apache.sysml.runtime.controlprogram.ParForProgramBlock.PExecMode used on opensource projects

Example 6 with PExecMode

use of org.apache.sysml.runtime.controlprogram.ParForProgramBlock.PExecMode in project incubator-systemml by apache.

the class OptimizerRuleBased method rewriteSetFusedDataPartitioningExecution.

// /////
// REWRITE set fused data partitioning / execution
// /
/**
 * This dedicated execution mode can only be applied if all of the
 * following conditions are true:
 * - Only cp instructions in the parfor body
 * - Only one partitioned input
 * - number of iterations is equal to number of partitions (nrow/ncol)
 * - partitioned matrix access via plain iteration variables (no composed expressions)
 *   (this ensures that each partition is exactly read once)
 * - no left indexing (since by default static task partitioning)
 *
 * Furthermore, it should be only chosen if we already decided for remote partitioning
 * and otherwise would create a large number of partition files.
 *
 * NOTE: We already respect the reducer memory budget for plan correctness. However,
 * we miss optimization potential if the reducer budget is larger than the mapper budget
 * (if we were not able to select REMOTE_MR as execution strategy wrt mapper budget)
 * TODO modify 'set exec strategy' and related rewrites for conditional data partitioning.
 *
 * @param pn internal representation of a plan alternative for program blocks and instructions
 * @param M ?
 * @param flagLIX ?
 * @param partitionedMatrices map of data partition formats
 * @param vars local variable map
 */
protected void rewriteSetFusedDataPartitioningExecution(OptNode pn, double M, boolean flagLIX, HashMap<String, PartitionFormat> partitionedMatrices, LocalVariableMap vars) {
    // assertions (warnings of corrupt optimizer decisions)
    if (pn.getNodeType() != NodeType.PARFOR)
        LOG.warn(getOptMode() + " OPT: Fused data partitioning and execution is only applicable for a ParFor node.");
    boolean apply = false;
    String partitioner = pn.getParam(ParamType.DATA_PARTITIONER);
    PDataPartitioner REMOTE_DP = OptimizerUtils.isSparkExecutionMode() ? PDataPartitioner.REMOTE_SPARK : PDataPartitioner.REMOTE_MR;
    PExecMode REMOTE_DPE = OptimizerUtils.isSparkExecutionMode() ? PExecMode.REMOTE_SPARK_DP : PExecMode.REMOTE_MR_DP;
    // try to merge MR data partitioning and MR exec
    if ((// fits into remote memory of reducers
    pn.getExecType() == ExecType.MR && M < _rm2 || // MR/SP EXEC and CP body
    pn.getExecType() == ExecType.SPARK) && partitioner != null && // MR/SP partitioning
    partitioner.equals(REMOTE_DP.toString()) && // only one partitioned matrix
    partitionedMatrices.size() == 1) {
        ParForProgramBlock pfpb = (ParForProgramBlock) OptTreeConverter.getAbstractPlanMapping().getMappedProg(pn.getID())[1];
        // partitioned matrix
        String moVarname = partitionedMatrices.keySet().iterator().next();
        PartitionFormat moDpf = partitionedMatrices.get(moVarname);
        MatrixObject mo = (MatrixObject) vars.get(moVarname);
        if (rIsAccessByIterationVariable(pn, moVarname, pfpb.getIterVar()) && ((moDpf == PartitionFormat.ROW_WISE && mo.getNumRows() == _N) || (moDpf == PartitionFormat.COLUMN_WISE && mo.getNumColumns() == _N) || (moDpf._dpf == PDataPartitionFormat.ROW_BLOCK_WISE_N && mo.getNumRows() <= _N * moDpf._N) || (moDpf._dpf == PDataPartitionFormat.COLUMN_BLOCK_WISE_N && mo.getNumColumns() <= _N * moDpf._N))) {
            int k = (int) Math.min(_N, _rk2);
            pn.addParam(ParamType.DATA_PARTITIONER, REMOTE_DPE.toString() + "(fused)");
            pn.setK(k);
            // set fused exec type
            pfpb.setExecMode(REMOTE_DPE);
            pfpb.setDataPartitioner(PDataPartitioner.NONE);
            pfpb.enableColocatedPartitionedMatrix(moVarname);
            pfpb.setDegreeOfParallelism(k);
            apply = true;
        }
    }
    LOG.debug(getOptMode() + " OPT: rewrite 'set fused data partitioning and execution' - result=" + apply);
}

Example 7 with PExecMode

use of org.apache.sysml.runtime.controlprogram.ParForProgramBlock.PExecMode in project systemml by apache.

the class OptimizerRuleBased method rewriteSetFusedDataPartitioningExecution.

// /////
// REWRITE set fused data partitioning / execution
// /
/**
 * This dedicated execution mode can only be applied if all of the
 * following conditions are true:
 * - Only cp instructions in the parfor body
 * - Only one partitioned input
 * - number of iterations is equal to number of partitions (nrow/ncol)
 * - partitioned matrix access via plain iteration variables (no composed expressions)
 *   (this ensures that each partition is exactly read once)
 * - no left indexing (since by default static task partitioning)
 *
 * Furthermore, it should be only chosen if we already decided for remote partitioning
 * and otherwise would create a large number of partition files.
 *
 * NOTE: We already respect the reducer memory budget for plan correctness. However,
 * we miss optimization potential if the reducer budget is larger than the mapper budget
 * (if we were not able to select REMOTE_MR as execution strategy wrt mapper budget)
 * TODO modify 'set exec strategy' and related rewrites for conditional data partitioning.
 *
 * @param pn internal representation of a plan alternative for program blocks and instructions
 * @param M ?
 * @param flagLIX ?
 * @param partitionedMatrices map of data partition formats
 * @param vars local variable map
 */
protected void rewriteSetFusedDataPartitioningExecution(OptNode pn, double M, boolean flagLIX, HashMap<String, PartitionFormat> partitionedMatrices, LocalVariableMap vars) {
    // assertions (warnings of corrupt optimizer decisions)
    if (pn.getNodeType() != NodeType.PARFOR)
        LOG.warn(getOptMode() + " OPT: Fused data partitioning and execution is only applicable for a ParFor node.");
    boolean apply = false;
    String partitioner = pn.getParam(ParamType.DATA_PARTITIONER);
    PDataPartitioner REMOTE_DP = OptimizerUtils.isSparkExecutionMode() ? PDataPartitioner.REMOTE_SPARK : PDataPartitioner.REMOTE_MR;
    PExecMode REMOTE_DPE = OptimizerUtils.isSparkExecutionMode() ? PExecMode.REMOTE_SPARK_DP : PExecMode.REMOTE_MR_DP;
    // try to merge MR data partitioning and MR exec
    if ((// fits into remote memory of reducers
    pn.getExecType() == ExecType.MR && M < _rm2 || // MR/SP EXEC and CP body
    pn.getExecType() == ExecType.SPARK) && partitioner != null && // MR/SP partitioning
    partitioner.equals(REMOTE_DP.toString()) && // only one partitioned matrix
    partitionedMatrices.size() == 1) {
        ParForProgramBlock pfpb = (ParForProgramBlock) OptTreeConverter.getAbstractPlanMapping().getMappedProg(pn.getID())[1];
        // partitioned matrix
        String moVarname = partitionedMatrices.keySet().iterator().next();
        PartitionFormat moDpf = partitionedMatrices.get(moVarname);
        MatrixObject mo = (MatrixObject) vars.get(moVarname);
        if (rIsAccessByIterationVariable(pn, moVarname, pfpb.getIterVar()) && ((moDpf == PartitionFormat.ROW_WISE && mo.getNumRows() == _N) || (moDpf == PartitionFormat.COLUMN_WISE && mo.getNumColumns() == _N) || (moDpf._dpf == PDataPartitionFormat.ROW_BLOCK_WISE_N && mo.getNumRows() <= _N * moDpf._N) || (moDpf._dpf == PDataPartitionFormat.COLUMN_BLOCK_WISE_N && mo.getNumColumns() <= _N * moDpf._N))) {
            int k = (int) Math.min(_N, _rk2);
            pn.addParam(ParamType.DATA_PARTITIONER, REMOTE_DPE.toString() + "(fused)");
            pn.setK(k);
            // set fused exec type
            pfpb.setExecMode(REMOTE_DPE);
            pfpb.setDataPartitioner(PDataPartitioner.NONE);
            pfpb.enableColocatedPartitionedMatrix(moVarname);
            pfpb.setDegreeOfParallelism(k);
            apply = true;
        }
    }
    LOG.debug(getOptMode() + " OPT: rewrite 'set fused data partitioning and execution' - result=" + apply);
}

Example 8 with PExecMode

use of org.apache.sysml.runtime.controlprogram.ParForProgramBlock.PExecMode in project systemml by apache.

the class OptimizerRuleBased method rewriteSetExecutionStategy.

// /////
// REWRITE set execution strategy
// /
protected boolean rewriteSetExecutionStategy(OptNode n, double M0, double M, double M2, double M3, boolean flagLIX) {
    boolean isCPOnly = n.isCPOnly();
    boolean isCPOnlyPossible = isCPOnly || isCPOnlyPossible(n, _rm);
    String datapartitioner = n.getParam(ParamType.DATA_PARTITIONER);
    ExecType REMOTE = getRemoteExecType();
    PDataPartitioner REMOTE_DP = OptimizerUtils.isSparkExecutionMode() ? PDataPartitioner.REMOTE_SPARK : PDataPartitioner.REMOTE_MR;
    // deciding on the execution strategy
    if (// allowed remote parfor execution
    ConfigurationManager.isParallelParFor() && (// Required: all inst already in cp and fit in remote mem
    (isCPOnly && M <= _rm) || // Required: all inst already in cp and fit partitioned in remote mem
    (isCPOnly && M3 <= _rm) || // Required: all inst forced to cp fit in remote mem
    (isCPOnlyPossible && M2 <= _rm))) {
        // at this point all required conditions for REMOTE_MR given, now its an opt decision
        // estimated local exploited par
        int cpk = (int) Math.min(_lk, Math.floor(_lm / M));
        // (the factor of 2 is to account for hyper-threading and in order prevent too eager remote parfor)
        if (// incl conditional partitioning
        2 * cpk < _lk && 2 * cpk < _N && 2 * cpk < _rk) {
            // remote parfor
            n.setExecType(REMOTE);
        } else // MR if problem is large enough and remote parallelism is larger than local
        if (_lk < _N && _lk < _rk && M <= _rm && isLargeProblem(n, M0)) {
            // remote parfor
            n.setExecType(REMOTE);
        } else // MR if MR operations in local, but CP only in remote (less overall MR jobs)
        if (!isCPOnly && isCPOnlyPossible) {
            // remote parfor
            n.setExecType(REMOTE);
        } else // MR if necessary for LIX rewrite (LIX true iff cp only and rm valid)
        if (flagLIX) {
            // remote parfor
            n.setExecType(REMOTE);
        } else // MR if remote data partitioning, because data will be distributed on all nodes
        if (datapartitioner != null && datapartitioner.equals(REMOTE_DP.toString()) && !InfrastructureAnalyzer.isLocalMode()) {
            // remote parfor
            n.setExecType(REMOTE);
        } else // otherwise CP
        {
            // local parfor
            n.setExecType(ExecType.CP);
        }
    } else // mr instructions in body, or rm too small
    {
        // local parfor
        n.setExecType(ExecType.CP);
    }
    // actual programblock modification
    long id = n.getID();
    ParForProgramBlock pfpb = (ParForProgramBlock) OptTreeConverter.getAbstractPlanMapping().getMappedProg(id)[1];
    PExecMode mode = n.getExecType().toParForExecMode();
    pfpb.setExecMode(mode);
    // decide if recompilation according to remote mem budget necessary
    boolean requiresRecompile = ((mode == PExecMode.REMOTE_MR || mode == PExecMode.REMOTE_SPARK) && !isCPOnly);
    _numEvaluatedPlans++;
    LOG.debug(getOptMode() + " OPT: rewrite 'set execution strategy' - result=" + mode + " (recompile=" + requiresRecompile + ")");
    return requiresRecompile;
}