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

Example 76 with ParForProgramBlock

use of org.apache.sysml.runtime.controlprogram.ParForProgramBlock 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 77 with ParForProgramBlock

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

the class OptimizerRuleBased method rewriteRemoveRecursiveParFor.

// /////
// REWRITE remove recursive parfor
// /
protected void rewriteRemoveRecursiveParFor(OptNode n, LocalVariableMap vars) {
    // num removed parfor
    int count = 0;
    // find recursive parfor
    HashSet<ParForProgramBlock> recPBs = new HashSet<>();
    rFindRecursiveParFor(n, recPBs, false);
    if (!recPBs.isEmpty()) {
        // unfold if necessary
        try {
            ParForProgramBlock pfpb = (ParForProgramBlock) OptTreeConverter.getAbstractPlanMapping().getMappedProg(n.getID())[1];
            if (recPBs.contains(pfpb))
                rFindAndUnfoldRecursiveFunction(n, pfpb, recPBs, vars);
        } catch (Exception ex) {
            throw new DMLRuntimeException(ex);
        }
        // remove recursive parfor (parfor to for)
        count = removeRecursiveParFor(n, recPBs);
    }
    _numEvaluatedPlans++;
    LOG.debug(getOptMode() + " OPT: rewrite 'remove recursive parfor' - result=" + recPBs.size() + "/" + count);
}

Example 78 with ParForProgramBlock

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

the class OptimizerRuleBased method rFindRecursiveParFor.

protected void rFindRecursiveParFor(OptNode n, HashSet<ParForProgramBlock> cand, boolean recContext) {
    // recursive invocation
    if (!n.isLeaf())
        for (OptNode c : n.getChilds()) {
            if (c.getNodeType() == NodeType.FUNCCALL && c.isRecursive())
                rFindRecursiveParFor(c, cand, true);
            else
                rFindRecursiveParFor(c, cand, recContext);
        }
    // add candidate program blocks
    if (recContext && n.getNodeType() == NodeType.PARFOR) {
        ParForProgramBlock pfpb = (ParForProgramBlock) OptTreeConverter.getAbstractPlanMapping().getMappedProg(n.getID())[1];
        cand.add(pfpb);
    }
}

Example 79 with ParForProgramBlock

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

the class OptimizerRuleBased method setTaskPartitioner.

protected void setTaskPartitioner(OptNode n, PTaskPartitioner partitioner) {
    long id = n.getID();
    // modify rtprog
    ParForProgramBlock pfpb = (ParForProgramBlock) OptTreeConverter.getAbstractPlanMapping().getMappedProg(id)[1];
    pfpb.setTaskPartitioner(partitioner);
    // modify plan
    n.addParam(ParamType.TASK_PARTITIONER, partitioner.toString());
    // handle specific case of LIX recompile
    boolean flagLIX = (partitioner == PTaskPartitioner.FACTORING_CMAX);
    if (flagLIX) {
        long maxc = n.getMaxC(_N);
        // used as constraint
        pfpb.setTaskSize(maxc);
        pfpb.disableJVMReuse();
        n.addParam(ParamType.TASK_SIZE, String.valueOf(maxc));
    }
    _numEvaluatedPlans++;
    LOG.debug(getOptMode() + " OPT: rewrite 'set task partitioner' - result=" + partitioner + ((flagLIX) ? "," + n.getParam(ParamType.TASK_SIZE) : ""));
}

Example 80 with ParForProgramBlock

use of org.apache.sysml.runtime.controlprogram.ParForProgramBlock 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;
}