@Override
public void processInstruction(ExecutionContext ec)
throws DMLRuntimeException, DMLUnsupportedOperationException {
SparkExecutionContext sec = (SparkExecutionContext) ec;
String rddVar = (_type == CacheType.LEFT) ? input2.getName() : input1.getName();
String bcastVar = (_type == CacheType.LEFT) ? input1.getName() : input2.getName();
MatrixCharacteristics mc = sec.getMatrixCharacteristics(output.getName());
long rlen =
sec.getScalarInput(_nrow.getName(), _nrow.getValueType(), _nrow.isLiteral()).getLongValue();
// get inputs
JavaPairRDD<MatrixIndexes, MatrixBlock> in1 = sec.getBinaryBlockRDDHandleForVariable(rddVar);
PartitionedBroadcastMatrix in2 = sec.getBroadcastForVariable(bcastVar);
// execute pmm instruction
JavaPairRDD<MatrixIndexes, MatrixBlock> out =
in1.flatMapToPair(new RDDPMMFunction(_type, in2, rlen, mc.getRowsPerBlock()));
out = RDDAggregateUtils.sumByKeyStable(out);
// put output RDD handle into symbol table
sec.setRDDHandleForVariable(output.getName(), out);
sec.addLineageRDD(output.getName(), rddVar);
sec.addLineageBroadcast(output.getName(), bcastVar);
// update output statistics if not inferred
updateBinaryMMOutputMatrixCharacteristics(sec, false);
}
@Override
protected long[] inferOutputCharacteristics(MemoTable memo) {
long[] ret = null;
Hop input = getInput().get(0);
MatrixCharacteristics mc = memo.getAllInputStats(input);
if (_direction == Direction.Col && mc.colsKnown()) ret = new long[] {1, mc.getCols(), -1};
else if (_direction == Direction.Row && mc.rowsKnown()) ret = new long[] {mc.getRows(), 1, -1};
return ret;
}
@Override
public void computeMemEstimate(MemoTable memo) {
// default behavior
super.computeMemEstimate(memo);
// try to infer via worstcase input statistics (for the case of dims known
// but nnz initially unknown)
MatrixCharacteristics mcM1 = memo.getAllInputStats(getInput().get(0));
if (dimsKnown() && mcM1.getNonZeros() >= 0) {
long lnnz = mcM1.getNonZeros(); // worst-case output nnz
double lOutMemEst = computeOutputMemEstimate(_dim1, _dim2, lnnz);
if (lOutMemEst < _outputMemEstimate) {
_outputMemEstimate = lOutMemEst;
_memEstimate = getInputOutputSize();
}
}
}
@Override
protected long[] inferOutputCharacteristics(MemoTable memo) {
long[] ret = null;
Hop input = getInput().get(0);
MatrixCharacteristics mc = memo.getAllInputStats(input);
if (mc.dimsKnown()) {
if (_op == OpOp1.ABS
|| _op == OpOp1.COS
|| _op == OpOp1.SIN
|| _op == OpOp1.TAN
|| _op == OpOp1.ACOS
|| _op == OpOp1.ASIN
|| _op == OpOp1.ATAN
|| _op == OpOp1.SQRT
|| _op == OpOp1.ROUND
|| _op == OpOp1.SPROP
|| _op == OpOp1.SELP) // sparsity preserving
{
ret = new long[] {mc.getRows(), mc.getCols(), mc.getNonZeros()};
} else ret = new long[] {mc.getRows(), mc.getCols(), -1};
}
return ret;
}
@Override
protected long[] inferOutputCharacteristics(MemoTable memo) {
long[] ret = null;
Hop input = getInput().get(0); // original matrix
MatrixCharacteristics mc = memo.getAllInputStats(input);
if (mc != null) {
long lnnz = mc.dimsKnown() ? Math.min(mc.getRows() * mc.getCols(), mc.getNonZeros()) : -1;
// worst-case is input size, but dense
ret = new long[] {mc.getRows(), mc.getCols(), lnnz};
// exploit column/row indexing information
if (_rowLowerEqualsUpper) ret[0] = 1;
if (_colLowerEqualsUpper) ret[1] = 1;
// infer tight block indexing size
Hop rl = getInput().get(1);
Hop ru = getInput().get(2);
Hop cl = getInput().get(3);
Hop cu = getInput().get(4);
if (isBlockIndexingExpression(rl, ru)) ret[0] = getBlockIndexingExpressionSize(rl, ru);
if (isBlockIndexingExpression(cl, cu)) ret[1] = getBlockIndexingExpressionSize(cl, cu);
}
return ret;
}
@Override
public void processInstruction(ExecutionContext ec)
throws DMLRuntimeException, DMLUnsupportedOperationException {
SparkExecutionContext sec = (SparkExecutionContext) ec;
MatrixCharacteristics mc = sec.getMatrixCharacteristics(input1.getName());
long rlen = mc.getRows();
int brlen = mc.getRowsPerBlock();
int bclen = mc.getColsPerBlock();
// get input
JavaPairRDD<MatrixIndexes, MatrixBlock> in =
sec.getBinaryBlockRDDHandleForVariable(input1.getName());
// execute unary aggregate (w/ implicit drop correction)
AggregateUnaryOperator auop = (AggregateUnaryOperator) _optr;
JavaPairRDD<MatrixIndexes, MatrixBlock> out =
in.mapToPair(new RDDCumAggFunction(auop, rlen, brlen, bclen));
out = RDDAggregateUtils.mergeByKey(out);
// put output handle in symbol table
sec.setRDDHandleForVariable(output.getName(), out);
sec.addLineageRDD(output.getName(), input1.getName());
}
/**
* @param type
* @param mcBc
* @return
*/
private static boolean requiresFlatMapFunction(CacheType type, MatrixCharacteristics mcBc) {
return (type == CacheType.LEFT && mcBc.getRows() > mcBc.getRowsPerBlock())
|| (type == CacheType.RIGHT && mcBc.getCols() > mcBc.getColsPerBlock());
}
/**
* @param mcIn
* @param type
* @return
*/
private static boolean preservesPartitioning(MatrixCharacteristics mcIn, CacheType type) {
if (type == CacheType.LEFT) return mcIn.dimsKnown() && mcIn.getRows() <= mcIn.getRowsPerBlock();
else // RIGHT
return mcIn.dimsKnown() && mcIn.getCols() <= mcIn.getColsPerBlock();
}
/**
* @param pfid
* @param program
* @param taskFile
* @param resultFile
* @param _enableCPCaching
* @param mode
* @param numMappers
* @param replication
* @return
* @throws DMLRuntimeException
*/
public static RemoteParForJobReturn runJob(
long pfid,
String program,
String taskFile,
String resultFile,
MatrixObject colocatedDPMatrixObj, // inputs
boolean enableCPCaching,
int numMappers,
int replication,
int max_retry,
long minMem,
boolean jvmReuse) // opt params
throws DMLRuntimeException {
RemoteParForJobReturn ret = null;
String jobname = "ParFor-EMR";
long t0 = DMLScript.STATISTICS ? System.nanoTime() : 0;
JobConf job;
job = new JobConf(RemoteParForMR.class);
job.setJobName(jobname + pfid);
// maintain dml script counters
Statistics.incrementNoOfCompiledMRJobs();
try {
/////
// configure the MR job
// set arbitrary CP program blocks that will perform in the mapper
MRJobConfiguration.setProgramBlocks(job, program);
// enable/disable caching
MRJobConfiguration.setParforCachingConfig(job, enableCPCaching);
// set mappers, reducers, combiners
job.setMapperClass(RemoteParWorkerMapper.class); // map-only
// set input format (one split per row, NLineInputFormat default N=1)
if (ParForProgramBlock.ALLOW_DATA_COLOCATION && colocatedDPMatrixObj != null) {
job.setInputFormat(RemoteParForColocatedNLineInputFormat.class);
MRJobConfiguration.setPartitioningFormat(job, colocatedDPMatrixObj.getPartitionFormat());
MatrixCharacteristics mc = colocatedDPMatrixObj.getMatrixCharacteristics();
MRJobConfiguration.setPartitioningBlockNumRows(job, mc.getRowsPerBlock());
MRJobConfiguration.setPartitioningBlockNumCols(job, mc.getColsPerBlock());
MRJobConfiguration.setPartitioningFilename(job, colocatedDPMatrixObj.getFileName());
} else // default case
{
job.setInputFormat(NLineInputFormat.class);
}
// set the input path and output path
FileInputFormat.setInputPaths(job, new Path(taskFile));
// set output format
job.setOutputFormat(SequenceFileOutputFormat.class);
// set output path
MapReduceTool.deleteFileIfExistOnHDFS(resultFile);
FileOutputFormat.setOutputPath(job, new Path(resultFile));
// set the output key, value schema
job.setMapOutputKeyClass(LongWritable.class);
job.setMapOutputValueClass(Text.class);
job.setOutputKeyClass(LongWritable.class);
job.setOutputValueClass(Text.class);
//////
// set optimization parameters
// set the number of mappers and reducers
job.setNumMapTasks(numMappers); // numMappers
job.setNumReduceTasks(0);
// job.setInt("mapred.map.tasks.maximum", 1); //system property
// job.setInt("mapred.tasktracker.tasks.maximum",1); //system property
// job.setInt("mapred.jobtracker.maxtasks.per.job",1); //system property
// use FLEX scheduler configuration properties
if (ParForProgramBlock.USE_FLEX_SCHEDULER_CONF) {
job.setInt("flex.priority", 0); // highest
job.setInt("flex.map.min", 0);
job.setInt("flex.map.max", numMappers);
job.setInt("flex.reduce.min", 0);
job.setInt("flex.reduce.max", numMappers);
}
// set jvm memory size (if require)
String memKey = "mapred.child.java.opts";
if (minMem > 0 && minMem > InfrastructureAnalyzer.extractMaxMemoryOpt(job.get(memKey))) {
InfrastructureAnalyzer.setMaxMemoryOpt(job, memKey, minMem);
LOG.warn("Forcing '" + memKey + "' to -Xmx" + minMem / (1024 * 1024) + "M.");
}
// disable automatic tasks timeouts and speculative task exec
job.setInt("mapred.task.timeout", 0);
job.setMapSpeculativeExecution(false);
// set up map/reduce memory configurations (if in AM context)
DMLConfig config = ConfigurationManager.getConfig();
DMLAppMasterUtils.setupMRJobRemoteMaxMemory(job, config);
// enables the reuse of JVMs (multiple tasks per MR task)
if (jvmReuse) job.setNumTasksToExecutePerJvm(-1); // unlimited
// set sort io buffer (reduce unnecessary large io buffer, guaranteed memory consumption)
job.setInt(MRConfigurationNames.MR_TASK_IO_SORT_MB, 8); // 8MB
// set the replication factor for the results
job.setInt("dfs.replication", replication);
// set the max number of retries per map task
// disabled job-level configuration to respect cluster configuration
// note: this refers to hadoop2, hence it never had effect on mr1
// job.setInt("mapreduce.map.maxattempts", max_retry);
// set unique working dir
MRJobConfiguration.setUniqueWorkingDir(job);
/////
// execute the MR job
RunningJob runjob = JobClient.runJob(job);
// Process different counters
Statistics.incrementNoOfExecutedMRJobs();
Group pgroup = runjob.getCounters().getGroup(ParForProgramBlock.PARFOR_COUNTER_GROUP_NAME);
int numTasks = (int) pgroup.getCounter(Stat.PARFOR_NUMTASKS.toString());
int numIters = (int) pgroup.getCounter(Stat.PARFOR_NUMITERS.toString());
if (DMLScript.STATISTICS && !InfrastructureAnalyzer.isLocalMode()) {
Statistics.incrementJITCompileTime(pgroup.getCounter(Stat.PARFOR_JITCOMPILE.toString()));
Statistics.incrementJVMgcCount(pgroup.getCounter(Stat.PARFOR_JVMGC_COUNT.toString()));
Statistics.incrementJVMgcTime(pgroup.getCounter(Stat.PARFOR_JVMGC_TIME.toString()));
Group cgroup =
runjob.getCounters().getGroup(CacheableData.CACHING_COUNTER_GROUP_NAME.toString());
CacheStatistics.incrementMemHits(
(int) cgroup.getCounter(CacheStatistics.Stat.CACHE_HITS_MEM.toString()));
CacheStatistics.incrementFSBuffHits(
(int) cgroup.getCounter(CacheStatistics.Stat.CACHE_HITS_FSBUFF.toString()));
CacheStatistics.incrementFSHits(
(int) cgroup.getCounter(CacheStatistics.Stat.CACHE_HITS_FS.toString()));
CacheStatistics.incrementHDFSHits(
(int) cgroup.getCounter(CacheStatistics.Stat.CACHE_HITS_HDFS.toString()));
CacheStatistics.incrementFSBuffWrites(
(int) cgroup.getCounter(CacheStatistics.Stat.CACHE_WRITES_FSBUFF.toString()));
CacheStatistics.incrementFSWrites(
(int) cgroup.getCounter(CacheStatistics.Stat.CACHE_WRITES_FS.toString()));
CacheStatistics.incrementHDFSWrites(
(int) cgroup.getCounter(CacheStatistics.Stat.CACHE_WRITES_HDFS.toString()));
CacheStatistics.incrementAcquireRTime(
cgroup.getCounter(CacheStatistics.Stat.CACHE_TIME_ACQR.toString()));
CacheStatistics.incrementAcquireMTime(
cgroup.getCounter(CacheStatistics.Stat.CACHE_TIME_ACQM.toString()));
CacheStatistics.incrementReleaseTime(
cgroup.getCounter(CacheStatistics.Stat.CACHE_TIME_RLS.toString()));
CacheStatistics.incrementExportTime(
cgroup.getCounter(CacheStatistics.Stat.CACHE_TIME_EXP.toString()));
}
// read all files of result variables and prepare for return
LocalVariableMap[] results = readResultFile(job, resultFile);
ret = new RemoteParForJobReturn(runjob.isSuccessful(), numTasks, numIters, results);
} catch (Exception ex) {
throw new DMLRuntimeException(ex);
} finally {
// remove created files
try {
MapReduceTool.deleteFileIfExistOnHDFS(new Path(taskFile), job);
MapReduceTool.deleteFileIfExistOnHDFS(new Path(resultFile), job);
} catch (IOException ex) {
throw new DMLRuntimeException(ex);
}
}
if (DMLScript.STATISTICS) {
long t1 = System.nanoTime();
Statistics.maintainCPHeavyHitters("MR-Job_" + jobname, t1 - t0);
}
return ret;
}
public void setMatrixCharacteristics(MatrixCharacteristics mcIn) {
_mcIn = mcIn;
_lastRowBlockIndex = (long) Math.ceil((double) _mcIn.getRows() / _mcIn.getRowsPerBlock());
}
public void computeMatrixCharacteristics(
MatrixCharacteristics mc1,
MatrixCharacteristics mc2,
MatrixCharacteristics mc3,
MatrixCharacteristics dimOut) {
QuaternaryOperator qop = (QuaternaryOperator) optr;
if (qop.wtype1 != null || qop.wtype4 != null) { // wsloss/wcemm
// output size independent of chain type (scalar)
dimOut.set(1, 1, mc1.getRowsPerBlock(), mc1.getColsPerBlock());
} else if (qop.wtype2 != null || qop.wtype5 != null) { // wsigmoid/wumm
// output size determined by main input
dimOut.set(mc1.getRows(), mc1.getCols(), mc1.getRowsPerBlock(), mc1.getColsPerBlock());
} else if (qop.wtype3 != null) { // wdivmm
// note: cannot directly consume mc2 or mc3 for redwdivmm because rep instruction changed
// the relevant dimensions; as a workaround the original dims are passed via nnz
boolean mapwdivmm = _cacheU && _cacheV;
long rank =
qop.wtype3.isLeft()
? mapwdivmm ? mc3.getCols() : mc3.getNonZeros()
: mapwdivmm ? mc2.getCols() : mc2.getNonZeros();
MatrixCharacteristics mcTmp =
qop.wtype3.computeOutputCharacteristics(mc1.getRows(), mc1.getCols(), rank);
dimOut.set(mcTmp.getRows(), mcTmp.getCols(), mc1.getRowsPerBlock(), mc1.getColsPerBlock());
}
}