use of org.apache.sysml.parser.ParForStatementBlock.ResultVar in project incubator-systemml by apache.
the class ParForProgramBlock method consolidateAndCheckResults.
private void consolidateAndCheckResults(ExecutionContext ec, long expIters, long expTasks, long numIters, long numTasks, LocalVariableMap[] results) {
Timing time = new Timing(true);
// result merge
if (checkParallelRemoteResultMerge()) {
// execute result merge in parallel for all result vars
int par = Math.min(_resultVars.size(), InfrastructureAnalyzer.getLocalParallelism());
if (InfrastructureAnalyzer.isLocalMode()) {
int parmem = (int) Math.floor(OptimizerUtils.getLocalMemBudget() / InfrastructureAnalyzer.getRemoteMaxMemorySortBuffer());
// reduce k if necessary
par = Math.min(par, Math.max(parmem, 1));
}
try {
// enqueue all result vars as tasks
LocalTaskQueue<ResultVar> q = new LocalTaskQueue<>();
for (ResultVar var : _resultVars) {
// foreach non-local write
if (// robustness scalars
ec.getVariable(var._name) instanceof MatrixObject)
q.enqueueTask(var);
}
q.closeInput();
// run result merge workers
ResultMergeWorker[] rmWorkers = new ResultMergeWorker[par];
for (int i = 0; i < par; i++) rmWorkers[i] = new ResultMergeWorker(q, results, ec);
for (// start all
int i = 0; // start all
i < par; // start all
i++) rmWorkers[i].start();
for (int i = 0; i < par; i++) {
// wait for all
rmWorkers[i].join();
if (!rmWorkers[i].finishedNoError())
throw new DMLRuntimeException("Error occured in parallel result merge worker.");
}
} catch (Exception ex) {
throw new DMLRuntimeException(ex);
}
} else {
// execute result merge sequentially for all result vars
for (// foreach non-local write
ResultVar var : // foreach non-local write
_resultVars) {
Data dat = ec.getVariable(var._name);
if (// robustness scalars
dat instanceof MatrixObject) {
MatrixObject out = (MatrixObject) dat;
MatrixObject[] in = new MatrixObject[results.length];
for (int i = 0; i < results.length; i++) in[i] = (MatrixObject) results[i].get(var._name);
String fname = constructResultMergeFileName();
ResultMerge rm = createResultMerge(_resultMerge, out, in, fname, var._isAccum, ec);
MatrixObject outNew = null;
if (USE_PARALLEL_RESULT_MERGE)
outNew = rm.executeParallelMerge(_numThreads);
else
outNew = rm.executeSerialMerge();
// cleanup existing var
Data exdata = ec.removeVariable(var._name);
if (exdata != null && exdata != outNew && exdata instanceof MatrixObject)
ec.cleanupCacheableData((MatrixObject) exdata);
// cleanup of intermediate result variables
cleanWorkerResultVariables(ec, out, in);
// set merged result variable
ec.setVariable(var._name, outNew);
}
}
}
// handle unscoped variables (vars created in parfor, but potentially used afterwards)
ParForStatementBlock sb = (ParForStatementBlock) getStatementBlock();
if (// sb might be null for nested parallelism
CREATE_UNSCOPED_RESULTVARS && sb != null && ec.getVariables() != null)
createEmptyUnscopedVariables(ec.getVariables(), sb);
// check expected counters
if (// consistency check
numTasks != expTasks || numIters != expIters)
throw new DMLRuntimeException("PARFOR: Number of executed tasks does not match the number of created tasks: tasks " + numTasks + "/" + expTasks + ", iters " + numIters + "/" + expIters + ".");
if (DMLScript.STATISTICS)
Statistics.incrementParForMergeTime((long) time.stop());
}
use of org.apache.sysml.parser.ParForStatementBlock.ResultVar in project incubator-systemml by apache.
the class ProgramConverter method parseResultVariables.
private static ArrayList<ResultVar> parseResultVariables(String in) {
ArrayList<ResultVar> ret = new ArrayList<>();
for (String var : parseStringArrayList(in)) {
boolean accum = var.endsWith("+");
ret.add(new ResultVar(accum ? var.substring(0, var.length() - 1) : var, accum));
}
return ret;
}
use of org.apache.sysml.parser.ParForStatementBlock.ResultVar in project incubator-systemml by apache.
the class RemoteParForUtils method exportResultVariables.
/**
* For remote Spark parfor workers. This is a simplified version compared to MR.
*
* @param workerID worker id
* @param vars local variable map
* @param resultVars list of result variables
* @return list of result variables
* @throws IOException if IOException occurs
*/
public static ArrayList<String> exportResultVariables(long workerID, LocalVariableMap vars, ArrayList<ResultVar> resultVars) throws IOException {
ArrayList<String> ret = new ArrayList<>();
// foreach result variables probe if export necessary
for (ResultVar rvar : resultVars) {
Data dat = vars.get(rvar._name);
// export output variable to HDFS (see RunMRJobs)
if (dat != null && dat.getDataType() == DataType.MATRIX) {
MatrixObject mo = (MatrixObject) dat;
if (mo.isDirty()) {
// export result var (iff actually modified in parfor)
mo.exportData();
// pass output vars (scalars by value, matrix by ref) to result
// (only if actually exported, hence in check for dirty, otherwise potential problems in result merge)
ret.add(ProgramConverter.serializeDataObject(rvar._name, mo));
}
}
}
return ret;
}
use of org.apache.sysml.parser.ParForStatementBlock.ResultVar in project incubator-systemml by apache.
the class RemoteParForUtils method exportResultVariables.
/**
* For remote MR parfor workers.
*
* @param workerID worker id
* @param vars local variable map
* @param resultVars list of result variables
* @param rvarFnames ?
* @param out output collectors
* @throws IOException if IOException occurs
*/
public static void exportResultVariables(long workerID, LocalVariableMap vars, ArrayList<ResultVar> resultVars, HashMap<String, String> rvarFnames, OutputCollector<Writable, Writable> out) throws IOException {
// create key and value for reuse
LongWritable okey = new LongWritable(workerID);
Text ovalue = new Text();
// foreach result variables probe if export necessary
for (ResultVar rvar : resultVars) {
Data dat = vars.get(rvar._name);
// export output variable to HDFS (see RunMRJobs)
if (dat != null && dat.getDataType() == DataType.MATRIX) {
MatrixObject mo = (MatrixObject) dat;
if (mo.isDirty()) {
if (ParForProgramBlock.ALLOW_REUSE_MR_PAR_WORKER && rvarFnames != null) {
String fname = rvarFnames.get(rvar._name);
if (fname != null)
mo.setFileName(fname);
// export result var (iff actually modified in parfor)
// note: this is equivalent to doing it in close (currently not required because 1 Task=1Map tasks, hence only one map invocation)
mo.exportData();
rvarFnames.put(rvar._name, mo.getFileName());
} else {
// export result var (iff actually modified in parfor)
// note: this is equivalent to doing it in close (currently not required because 1 Task=1Map tasks, hence only one map invocation)
mo.exportData();
}
// pass output vars (scalars by value, matrix by ref) to result
// (only if actually exported, hence in check for dirty, otherwise potential problems in result merge)
String datStr = ProgramConverter.serializeDataObject(rvar._name, mo);
ovalue.set(datStr);
out.collect(okey, ovalue);
}
}
}
}
Aggregations