use of org.apache.sysml.runtime.controlprogram.ParForProgramBlock.PartitionFormat in project incubator-systemml by apache.
the class ParForStatementBlock method determineAccessPattern.
private PartitionFormat determineAccessPattern(IndexedIdentifier dat) {
boolean isSpark = OptimizerUtils.isSparkExecutionMode();
int blksz = ConfigurationManager.getBlocksize();
PartitionFormat dpf = null;
// 1) get all bounds expressions for index access
Expression rowL = dat.getRowLowerBound();
Expression rowU = dat.getRowUpperBound();
Expression colL = dat.getColLowerBound();
Expression colU = dat.getColUpperBound();
boolean allRows = (rowL == null && rowU == null);
boolean allCols = (colL == null && colU == null);
try {
// COLUMN_WISE if all rows and access to single column
if (allRows && colL != null && colL.equals(colU)) {
dpf = PartitionFormat.COLUMN_WISE;
} else // ROW_WISE if all cols and access to single row
if (allCols && rowL != null && rowL.equals(rowU)) {
dpf = PartitionFormat.ROW_WISE;
} else // COLUMN_BLOCK_WISE
if (isSpark && allRows && colL != colU) {
LinearFunction l1 = getLinearFunction(colL, true);
LinearFunction l2 = getLinearFunction(colU, true);
dpf = !isAlignedBlocking(l1, l2, blksz) ? PartitionFormat.NONE : new PartitionFormat(PDataPartitionFormat.COLUMN_BLOCK_WISE_N, (int) l1._b[0]);
} else // ROW_BLOCK_WISE
if (isSpark && allCols && rowL != rowU) {
LinearFunction l1 = getLinearFunction(rowL, true);
LinearFunction l2 = getLinearFunction(rowU, true);
dpf = !isAlignedBlocking(l1, l2, blksz) ? PartitionFormat.NONE : new PartitionFormat(PDataPartitionFormat.ROW_BLOCK_WISE_N, (int) l1._b[0]);
} else
// NONE otherwise (conservative)
dpf = PartitionFormat.NONE;
} catch (Exception ex) {
throw new RuntimeException(ex);
}
return dpf;
}
use of org.apache.sysml.runtime.controlprogram.ParForProgramBlock.PartitionFormat in project incubator-systemml by apache.
the class ParForStatementBlock method determineDataPartitionFormat.
/**
* Determines the PDataPartitioningFormat for read-only parent variables according
* to the access pattern of that variable within the parfor statement block.
* Row-wise or column wise partitioning is only suggested if we see pure row-wise or
* column-wise access patterns.
*
* @param var variables
* @return partition format
*/
public PartitionFormat determineDataPartitionFormat(String var) {
PartitionFormat dpf = null;
List<PartitionFormat> dpfc = new LinkedList<>();
try {
// determine partitioning candidates
ParForStatement dpfs = (ParForStatement) _statements.get(0);
rDeterminePartitioningCandidates(var, dpfs.getBody(), dpfc);
// determine final solution
for (PartitionFormat tmp : dpfc) dpf = // if no consensus
(dpf != null && !dpf.equals(tmp)) ? PartitionFormat.NONE : tmp;
if (dpf == null)
dpf = PartitionFormat.NONE;
} catch (LanguageException e) {
LOG.trace("Unable to determine partitioning candidates.", e);
dpf = PartitionFormat.NONE;
}
return dpf;
}
use of org.apache.sysml.runtime.controlprogram.ParForProgramBlock.PartitionFormat in project incubator-systemml by apache.
the class RemoteParForColocatedNLineInputFormat method getSplits.
@Override
public InputSplit[] getSplits(JobConf job, int numSplits) throws IOException {
InputSplit[] tmp = super.getSplits(job, numSplits);
// get partitioning information
MatrixCharacteristics mc = MRJobConfiguration.getPartitionedMatrixSize(job);
PDataPartitionFormat dpf = MRJobConfiguration.getPartitioningFormat(job);
PartitionFormat pf = new PartitionFormat(dpf, -1);
int blen = (int) (pf.isRowwise() ? pf.getNumRows(mc) : pf.getNumColumns(mc));
String fname = MRJobConfiguration.getPartitioningFilename(job);
// create wrapper splits
InputSplit[] ret = new InputSplit[tmp.length];
for (int i = 0; i < tmp.length; i++) {
// check for robustness of subsequent cast
if (tmp[i] instanceof FileSplit)
ret[i] = new RemoteParForColocatedFileSplit((FileSplit) tmp[i], fname, blen);
else
ret[i] = tmp[i];
}
return ret;
}
use of org.apache.sysml.runtime.controlprogram.ParForProgramBlock.PartitionFormat in project incubator-systemml by apache.
the class OptimizerConstrained method rewriteSetFusedDataPartitioningExecution.
// /////
// REWRITE set fused data partitioning / execution
// /
protected void rewriteSetFusedDataPartitioningExecution(OptNode pn, double M, boolean flagLIX, HashMap<String, PartitionFormat> partitionedMatrices, LocalVariableMap vars, PExecMode emode) {
if (emode == PExecMode.REMOTE_MR_DP || emode == PExecMode.REMOTE_SPARK_DP) {
ParForProgramBlock pfpb = (ParForProgramBlock) OptTreeConverter.getAbstractPlanMapping().getMappedProg(pn.getID())[1];
// partitioned matrix
if (partitionedMatrices.size() <= 0) {
LOG.debug(getOptMode() + " OPT: unable to force 'set fused data partitioning and execution' - result=" + false);
return;
}
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);
if (emode == PExecMode.REMOTE_MR_DP) {
pn.addParam(ParamType.DATA_PARTITIONER, "REMOTE_MR(fused)");
// set fused exec type
pfpb.setExecMode(PExecMode.REMOTE_MR_DP);
} else {
pn.addParam(ParamType.DATA_PARTITIONER, "REMOTE_SPARK(fused)");
// set fused exec type
pfpb.setExecMode(PExecMode.REMOTE_SPARK_DP);
}
pn.setK(k);
pfpb.setDataPartitioner(PDataPartitioner.NONE);
pfpb.enableColocatedPartitionedMatrix(moVarname);
pfpb.setDegreeOfParallelism(k);
}
LOG.debug(getOptMode() + " OPT: force 'set fused data partitioning and execution' - result=" + true);
} else
super.rewriteSetFusedDataPartitioningExecution(pn, M, flagLIX, partitionedMatrices, vars);
}
use of org.apache.sysml.runtime.controlprogram.ParForProgramBlock.PartitionFormat 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);
}
Aggregations