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

Example 41 with ParForProgramBlock

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

the class OptimizerRuleBased method rewriteSetTranposeSparseVectorOperations.

// /////
// REWRITE transpose sparse vector operations
// /
protected void rewriteSetTranposeSparseVectorOperations(OptNode pn, HashMap<String, PartitionFormat> partitionedMatrices, LocalVariableMap vars) {
    // assertions (warnings of corrupt optimizer decisions)
    if (pn.getNodeType() != NodeType.PARFOR)
        LOG.warn(getOptMode() + " OPT: Transpose sparse vector operations is only applicable for a ParFor node.");
    boolean apply = false;
    ParForProgramBlock pfpb = (ParForProgramBlock) OptTreeConverter.getAbstractPlanMapping().getMappedProg(pn.getID())[1];
    if (pfpb.getExecMode() == PExecMode.REMOTE_MR_DP && // general applicable
    partitionedMatrices.size() == 1) {
        String moVarname = partitionedMatrices.keySet().iterator().next();
        PartitionFormat moDpf = partitionedMatrices.get(moVarname);
        Data dat = vars.get(moVarname);
        if (dat != null && dat instanceof MatrixObject && moDpf == PartitionFormat.COLUMN_WISE && // check for sparse matrix
        ((MatrixObject) dat).getSparsity() <= MatrixBlock.SPARSITY_TURN_POINT && // tranpose-safe
        rIsTransposeSafePartition(pn, moVarname)) {
            pfpb.setTransposeSparseColumnVector(true);
            apply = true;
        }
    }
    LOG.debug(getOptMode() + " OPT: rewrite 'set transpose sparse vector operations' - result=" + apply);
}

Example 42 with ParForProgramBlock

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

the class OptimizerRuleBased method rewriteSetRecompileMemoryBudget.

// /////
// REWRITE set recompile memory budget
// /
protected void rewriteSetRecompileMemoryBudget(OptNode n) {
    double newLocalMem = _lm;
    // check et because recompilation only happens at the master node
    if (n.getExecType() == ExecType.CP) {
        // compute local recompile memory budget
        int par = n.getTotalK();
        newLocalMem = _lm / par;
        // modify runtime plan
        ParForProgramBlock pfpb = (ParForProgramBlock) OptTreeConverter.getAbstractPlanMapping().getMappedProg(n.getID())[1];
        pfpb.setRecompileMemoryBudget(newLocalMem);
    }
    _numEvaluatedPlans++;
    LOG.debug(getOptMode() + " OPT: rewrite 'set recompile memory budget' - result=" + toMB(newLocalMem));
}

Example 43 with ParForProgramBlock

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

the class OptimizerRuleBased method rFindAndUnfoldRecursiveFunction.

protected void rFindAndUnfoldRecursiveFunction(OptNode n, ParForProgramBlock parfor, HashSet<ParForProgramBlock> recPBs, LocalVariableMap vars) {
    // unfold if found
    if (n.getNodeType() == NodeType.FUNCCALL && n.isRecursive()) {
        boolean exists = rContainsNode(n, parfor);
        if (exists) {
            String fnameKey = n.getParam(ParamType.OPSTRING);
            String[] names = fnameKey.split(Program.KEY_DELIM);
            String fnamespace = names[0];
            String fname = names[1];
            String fnameNew = FUNCTION_UNFOLD_NAMEPREFIX + fname;
            // unfold function
            FunctionOp fop = (FunctionOp) OptTreeConverter.getAbstractPlanMapping().getMappedHop(n.getID());
            Program prog = parfor.getProgram();
            DMLProgram dmlprog = parfor.getStatementBlock().getDMLProg();
            FunctionProgramBlock fpb = prog.getFunctionProgramBlock(fnamespace, fname);
            FunctionProgramBlock copyfpb = ProgramConverter.createDeepCopyFunctionProgramBlock(fpb, new HashSet<String>(), new HashSet<String>());
            prog.addFunctionProgramBlock(fnamespace, fnameNew, copyfpb);
            dmlprog.addFunctionStatementBlock(fnamespace, fnameNew, (FunctionStatementBlock) copyfpb.getStatementBlock());
            // replace function names in old subtree (link to new function)
            rReplaceFunctionNames(n, fname, fnameNew);
            // recreate sub opttree
            String fnameNewKey = fnamespace + Program.KEY_DELIM + fnameNew;
            OptNode nNew = new OptNode(NodeType.FUNCCALL);
            OptTreeConverter.getAbstractPlanMapping().putHopMapping(fop, nNew);
            nNew.setExecType(ExecType.CP);
            nNew.addParam(ParamType.OPSTRING, fnameNewKey);
            long parentID = OptTreeConverter.getAbstractPlanMapping().getMappedParentID(n.getID());
            OptTreeConverter.getAbstractPlanMapping().getOptNode(parentID).exchangeChild(n, nNew);
            HashSet<String> memo = new HashSet<>();
            // required if functionop not shared (because not replaced yet)
            memo.add(fnameKey);
            // requied if functionop shared (indirectly replaced)
            memo.add(fnameNewKey);
            for (int i = 0; i < copyfpb.getChildBlocks().size(); /*&& i<len*/
            i++) {
                ProgramBlock lpb = copyfpb.getChildBlocks().get(i);
                StatementBlock lsb = lpb.getStatementBlock();
                nNew.addChild(OptTreeConverter.rCreateAbstractOptNode(lsb, lpb, vars, false, memo));
            }
            // compute delta for recPB set (use for removing parfor)
            recPBs.removeAll(rGetAllParForPBs(n, new HashSet<ParForProgramBlock>()));
            recPBs.addAll(rGetAllParForPBs(nNew, new HashSet<ParForProgramBlock>()));
            // replace function names in new subtree (recursive link to new function)
            rReplaceFunctionNames(nNew, fname, fnameNew);
        }
        return;
    }
    // recursive invocation (only for non-recursive functions)
    if (!n.isLeaf())
        for (OptNode c : n.getChilds()) rFindAndUnfoldRecursiveFunction(c, parfor, recPBs, vars);
}

Example 44 with ParForProgramBlock

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

the class OptimizerRuleBased method rAssignRemainingParallelism.

protected void rAssignRemainingParallelism(OptNode n, int parforK, int opsK) {
    ArrayList<OptNode> childs = n.getChilds();
    if (childs != null) {
        boolean recompileSB = false;
        for (OptNode c : childs) {
            if (c.getNodeType() == NodeType.PARFOR) {
                // constrain max parfor parallelism by problem size
                int tmpN = Integer.parseInt(c.getParam(ParamType.NUM_ITERATIONS));
                int tmpK = (tmpN < parforK) ? tmpN : parforK;
                // set parfor degree of parallelism
                long id = c.getID();
                c.setK(tmpK);
                ParForProgramBlock pfpb = (ParForProgramBlock) OptTreeConverter.getAbstractPlanMapping().getMappedProg(id)[1];
                pfpb.setDegreeOfParallelism(tmpK);
                // distribute remaining parallelism
                int remainParforK = getRemainingParallelismParFor(parforK, tmpK);
                int remainOpsK = getRemainingParallelismOps(opsK, tmpK);
                rAssignRemainingParallelism(c, remainParforK, remainOpsK);
            } else if (c.getNodeType() == NodeType.HOP) {
                // set degree of parallelism for multi-threaded leaf nodes
                Hop h = OptTreeConverter.getAbstractPlanMapping().getMappedHop(c.getID());
                if (ConfigurationManager.isParallelMatrixOperations() && // abop, datagenop, qop, paramop
                h instanceof MultiThreadedHop && !(// only paramop-grpagg
                h instanceof ParameterizedBuiltinOp && !HopRewriteUtils.isValidOp(((ParameterizedBuiltinOp) h).getOp(), ParamBuiltinOp.GROUPEDAGG, ParamBuiltinOp.REXPAND)) && !(// only unaryop-cumulativeagg
                h instanceof UnaryOp && !((UnaryOp) h).isCumulativeUnaryOperation()) && !(// only reorgop-transpose
                h instanceof ReorgOp && ((ReorgOp) h).getOp() != ReOrgOp.TRANSPOSE)) {
                    MultiThreadedHop mhop = (MultiThreadedHop) h;
                    // set max constraint in hop
                    mhop.setMaxNumThreads(opsK);
                    // set optnode k (for explain)
                    c.setK(opsK);
                    // need to recompile SB, if changed constraint
                    recompileSB = true;
                } else // for all other multi-threaded hops set k=1 to simply debugging
                if (h instanceof MultiThreadedHop) {
                    MultiThreadedHop mhop = (MultiThreadedHop) h;
                    // set max constraint in hop
                    mhop.setMaxNumThreads(1);
                    // set optnode k (for explain)
                    c.setK(1);
                }
            } else
                rAssignRemainingParallelism(c, parforK, opsK);
        }
        // recompile statement block if required
        if (recompileSB) {
            try {
                // guaranteed to be a last-level block (see hop change)
                ProgramBlock pb = (ProgramBlock) OptTreeConverter.getAbstractPlanMapping().getMappedProg(n.getID())[1];
                Recompiler.recompileProgramBlockInstructions(pb);
            } catch (Exception ex) {
                throw new DMLRuntimeException(ex);
            }
        }
    }
}

Example 45 with ParForProgramBlock

use of org.apache.sysml.runtime.controlprogram.ParForProgramBlock in project incubator-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;
}