/*
* Licensed to the Apache Software Foundation (ASF) under one
* or more contributor license agreements. See the NOTICE file
* distributed with this work for additional information
* regarding copyright ownership. The ASF licenses this file
* to you under the Apache License, Version 2.0 (the
* "License"); you may not use this file except in compliance
* with the License. You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing,
* software distributed under the License is distributed on an
* "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
* KIND, either express or implied. See the License for the
* specific language governing permissions and limitations
* under the License.
*/
package org.apache.sysml.lops.runtime;
import java.io.IOException;
import org.apache.hadoop.fs.Path;
import org.apache.wink.json4j.JSONException;
import org.apache.sysml.api.DMLScript;
import org.apache.sysml.api.DMLScript.RUNTIME_PLATFORM;
import org.apache.sysml.conf.ConfigurationManager;
import org.apache.sysml.conf.DMLConfig;
import org.apache.sysml.hops.OptimizerUtils;
import org.apache.sysml.hops.recompile.Recompiler;
import org.apache.sysml.lops.Lop;
import org.apache.sysml.lops.compile.JobType;
import org.apache.sysml.parser.Expression.DataType;
import org.apache.sysml.parser.Expression.ValueType;
import org.apache.sysml.runtime.DMLRuntimeException;
import org.apache.sysml.runtime.controlprogram.LocalVariableMap;
import org.apache.sysml.runtime.controlprogram.caching.MatrixObject;
import org.apache.sysml.runtime.controlprogram.context.ExecutionContext;
import org.apache.sysml.runtime.controlprogram.parfor.mqo.RuntimePiggybacking;
import org.apache.sysml.runtime.instructions.MRInstructionParser;
import org.apache.sysml.runtime.instructions.MRJobInstruction;
import org.apache.sysml.runtime.instructions.cp.Data;
import org.apache.sysml.runtime.instructions.cp.ScalarObject;
import org.apache.sysml.runtime.instructions.mr.DataGenMRInstruction;
import org.apache.sysml.runtime.instructions.mr.RandInstruction;
import org.apache.sysml.runtime.instructions.mr.ReblockInstruction;
import org.apache.sysml.runtime.instructions.mr.SeqInstruction;
import org.apache.sysml.runtime.io.MatrixWriter;
import org.apache.sysml.runtime.io.MatrixWriterFactory;
import org.apache.sysml.runtime.matrix.CMCOVMR;
import org.apache.sysml.runtime.matrix.CSVReblockMR;
import org.apache.sysml.runtime.matrix.CombineMR;
import org.apache.sysml.runtime.matrix.DataGenMR;
import org.apache.sysml.runtime.matrix.DataPartitionMR;
import org.apache.sysml.runtime.matrix.GMR;
import org.apache.sysml.runtime.matrix.GroupedAggMR;
import org.apache.sysml.runtime.matrix.JobReturn;
import org.apache.sysml.runtime.matrix.MMCJMR;
import org.apache.sysml.runtime.matrix.MMRJMR;
import org.apache.sysml.runtime.matrix.MatrixCharacteristics;
import org.apache.sysml.runtime.matrix.MatrixDimensionsMetaData;
import org.apache.sysml.runtime.matrix.MatrixFormatMetaData;
import org.apache.sysml.runtime.matrix.ReblockMR;
import org.apache.sysml.runtime.matrix.SortMR;
import org.apache.sysml.runtime.matrix.WriteCSVMR;
import org.apache.sysml.runtime.matrix.data.LibMatrixDatagen;
import org.apache.sysml.runtime.matrix.data.MatrixBlock;
import org.apache.sysml.runtime.matrix.data.OutputInfo;
import org.apache.sysml.runtime.matrix.data.RandomMatrixGenerator;
import org.apache.sysml.runtime.transform.DataTransform;
import org.apache.sysml.runtime.util.MapReduceTool;
import org.apache.sysml.utils.Statistics;
public class RunMRJobs
{
/**
* Wrapper for submitting MR job instructions incl preparation and actual submission.
* The preparation includes (1) pulling stats out of symbol table and populating the
* instruction, (2) instruction patching, and (3) export of in-memory matrices if
* required.
*
* Furthermore, this wrapper also provides a hook for runtime piggybacking to intercept
* concurrent job submissions in order to collect and merge instructions.
*
* @param inst instruction
* @param ec execution context
* @return job status
* @throws DMLRuntimeException if DMLRuntimeException occurs
*/
public static JobReturn prepareAndSubmitJob( MRJobInstruction inst, ExecutionContext ec )
throws DMLRuntimeException
{
// Obtain references to all input matrices
MatrixObject[] inputMatrices = inst.extractInputMatrices(ec);
// export dirty matrices to HDFS
// note: for REBLOCK postponed until we know if necessary
if( !(inst.getJobType() == JobType.REBLOCK) )
{
//export matrices
for(MatrixObject m : inputMatrices) {
if ( m.isDirty() || m.getRDDHandle()!=null )
m.exportData();
}
//check input files
checkEmptyInputs( inst, inputMatrices );
}
// Obtain references to all output matrices
inst.extractOutputMatrices(ec);
// obtain original state
String rdInst = inst.getIv_randInstructions();
String rrInst = inst.getIv_recordReaderInstructions();
String mapInst = inst.getIv_instructionsInMapper();
String shuffleInst = inst.getIv_shuffleInstructions();
String aggInst = inst.getIv_aggInstructions();
String otherInst = inst.getIv_otherInstructions();
// variable patching (replace placeholders with variables)
inst.setIv_randInstructions(updateLabels(rdInst, ec.getVariables()));
inst.setIv_recordReaderInstructions(updateLabels(rrInst, ec.getVariables()));
inst.setIv_instructionsInMapper(updateLabels(mapInst, ec.getVariables()));
inst.setIv_shuffleInstructions(updateLabels(shuffleInst, ec.getVariables()));
inst.setIv_aggInstructions(updateLabels(aggInst, ec.getVariables()));
inst.setIv_otherInstructions(updateLabels(otherInst, ec.getVariables()));
// runtime piggybacking if applicable
JobReturn ret = null;
if( OptimizerUtils.ALLOW_RUNTIME_PIGGYBACKING
&& RuntimePiggybacking.isActive()
&& RuntimePiggybacking.isSupportedJobType(inst.getJobType()) )
{
ret = RuntimePiggybacking.submitJob( inst );
}
else
ret = submitJob( inst );
// reset original state
inst.setIv_randInstructions(rdInst);
inst.setIv_recordReaderInstructions(rrInst);
inst.setIv_instructionsInMapper(mapInst);
inst.setIv_shuffleInstructions(shuffleInst);
inst.setIv_aggInstructions(aggInst);
inst.setIv_otherInstructions(otherInst);
return ret;
}
/**
* Submits an MR job instruction, without modifying any state of that instruction.
*
* @param inst instruction
* @return job status
* @throws DMLRuntimeException if DMLRuntimeException occurs
*/
public static JobReturn submitJob(MRJobInstruction inst )
throws DMLRuntimeException
{
JobReturn ret = new JobReturn();
MatrixObject[] inputMatrices = inst.getInputMatrices();
MatrixObject[] outputMatrices = inst.getOutputMatrices();
boolean execCP = false;
// Spawn MapReduce Jobs
try {
// replace all placeholders in all instructions with appropriate values
String rdInst = inst.getIv_randInstructions();
String rrInst = inst.getIv_recordReaderInstructions();
String mapInst = inst.getIv_instructionsInMapper();
String shuffleInst = inst.getIv_shuffleInstructions();
String aggInst = inst.getIv_aggInstructions();
String otherInst = inst.getIv_otherInstructions();
boolean jvmReuse = ConfigurationManager.getDMLConfig().getBooleanValue(DMLConfig.JVM_REUSE);
switch(inst.getJobType()) {
case GMR:
case GMRCELL:
ret = GMR.runJob(inst, inst.getInputs(), inst.getInputInfos(),
inst.getRlens(), inst.getClens(), inst.getBrlens(), inst.getBclens(),
inst.getPartitioned(), inst.getPformats(), inst.getPsizes(),
rrInst, mapInst, aggInst, otherInst,
inst.getIv_numReducers(), inst.getIv_replication(), jvmReuse, inst.getIv_resultIndices(), inst.getDimsUnknownFilePrefix(),
inst.getOutputs(), inst.getOutputInfos() );
break;
case DATAGEN:
if( ConfigurationManager.isDynamicRecompilation()
&& OptimizerUtils.ALLOW_RAND_JOB_RECOMPILE
&& DMLScript.rtplatform != RUNTIME_PLATFORM.HADOOP
&& Recompiler.checkCPDataGen( inst, rdInst ) )
{
ret = executeInMemoryDataGenOperations(inst, rdInst, outputMatrices);
Statistics.decrementNoOfExecutedMRJobs();
execCP = true;
}
else
{
ret = DataGenMR.runJob(inst,
rdInst.split(Lop.INSTRUCTION_DELIMITOR), mapInst, aggInst, otherInst,
inst.getIv_numReducers(), inst.getIv_replication(), inst.getIv_resultIndices(), inst.getDimsUnknownFilePrefix(),
inst.getOutputs(), inst.getOutputInfos());
}
break;
case CM_COV:
ret = CMCOVMR.runJob(inst, inst.getInputs(), inst.getInputInfos(),
inst.getRlens(), inst.getClens(), inst.getBrlens(), inst.getBclens(),
mapInst, shuffleInst,
inst.getIv_numReducers(), inst.getIv_replication(), inst.getIv_resultIndices(),
inst.getOutputs(), inst.getOutputInfos() );
break;
case GROUPED_AGG:
ret = GroupedAggMR.runJob(inst, inst.getInputs(), inst.getInputInfos(),
inst.getRlens(), inst.getClens(), inst.getBrlens(), inst.getBclens(),
shuffleInst, otherInst,
inst.getIv_numReducers(), inst.getIv_replication(), inst.getIv_resultIndices(), inst.getDimsUnknownFilePrefix(),
inst.getOutputs(), inst.getOutputInfos() );
break;
case REBLOCK:
case CSV_REBLOCK:
if( ConfigurationManager.isDynamicRecompilation()
&& DMLScript.rtplatform != RUNTIME_PLATFORM.HADOOP
&& Recompiler.checkCPReblock( inst, inputMatrices ) )
{
ret = executeInMemoryReblockOperations(inst, shuffleInst, inputMatrices, outputMatrices);
Statistics.decrementNoOfExecutedMRJobs();
execCP = true;
}
else
{
// export dirty matrices to HDFS (initially deferred)
for(MatrixObject m : inputMatrices) {
if ( m.isDirty() )
m.exportData();
}
checkEmptyInputs( inst, inputMatrices );
if ( inst.getJobType() == JobType.REBLOCK ) {
ret = ReblockMR.runJob(inst, inst.getInputs(), inst.getInputInfos(),
inst.getRlens(), inst.getClens(), inst.getBrlens(), inst.getBclens(), getNNZ(inputMatrices),
mapInst, shuffleInst, otherInst,
inst.getIv_numReducers(), inst.getIv_replication(), jvmReuse, inst.getIv_resultIndices(),
inst.getOutputs(), inst.getOutputInfos() );
}
else if( inst.getJobType() == JobType.CSV_REBLOCK ) {
ret = CSVReblockMR.runJob(inst, inst.getInputs(), inst.getInputInfos(),
inst.getRlens(), inst.getClens(), inst.getBrlens(), inst.getBclens(),
shuffleInst, otherInst,
inst.getIv_numReducers(), inst.getIv_replication(), inst.getIv_resultIndices(),
inst.getOutputs(), inst.getOutputInfos() );
}
}
break;
case CSV_WRITE:
ret = WriteCSVMR.runJob(inst, inst.getInputs(), inst.getInputInfos(),
inst.getRlens(), inst.getClens(), inst.getBclens(), inst.getBclens(), shuffleInst,
inst.getIv_numReducers(), inst.getIv_replication(), inst.getIv_resultIndices(), inst.getOutputs());
break;
case MMCJ:
ret = MMCJMR.runJob(inst, inst.getInputs(), inst.getInputInfos(),
inst.getRlens(), inst.getClens(), inst.getBrlens(), inst.getBclens(),
mapInst, aggInst, shuffleInst,
inst.getIv_numReducers(), inst.getIv_replication(),
inst.getOutputs()[0], inst.getOutputInfos()[0] );
break;
case MMRJ:
ret = MMRJMR.runJob(inst, inst.getInputs(), inst.getInputInfos(),
inst.getRlens(), inst.getClens(), inst.getBrlens(), inst.getBclens(),
mapInst, aggInst, shuffleInst, otherInst,
inst.getIv_numReducers(), inst.getIv_replication(), inst.getIv_resultIndices(),
inst.getOutputs(), inst.getOutputInfos() );
break;
case SORT:
boolean weightsflag = true;
if ( !mapInst.equalsIgnoreCase("") )
weightsflag = false;
ret = SortMR.runJob(inst, inst.getInputs()[0], inst.getInputInfos()[0],
inst.getRlens()[0], inst.getClens()[0], inst.getBrlens()[0], inst.getBclens()[0],
mapInst, shuffleInst,
inst.getIv_numReducers(), inst.getIv_replication(),
inst.getOutputs()[0], inst.getOutputInfos()[0], weightsflag );
break;
case COMBINE:
ret = CombineMR.runJob(inst, inst.getInputs(), inst.getInputInfos(),
inst.getRlens(), inst.getClens(), inst.getBrlens(), inst.getBclens(),
shuffleInst,
inst.getIv_numReducers(), inst.getIv_replication(), inst.getIv_resultIndices(),
inst.getOutputs(), inst.getOutputInfos() );
break;
case DATA_PARTITION:
ret = DataPartitionMR.runJob(inst, inputMatrices, shuffleInst, inst.getIv_resultIndices(), outputMatrices, inst.getIv_numReducers(), inst.getIv_replication());
break;
case TRANSFORM:
if( ConfigurationManager.isDynamicRecompilation()
&& OptimizerUtils.ALLOW_TRANSFORM_RECOMPILE
&& DMLScript.rtplatform != RUNTIME_PLATFORM.HADOOP
&& Recompiler.checkCPTransform( inst, inputMatrices ) )
{
// transform the data and generate output in CSV format
ret = executeInMemoryTransform(inst, inputMatrices, outputMatrices);
Statistics.decrementNoOfExecutedMRJobs();
execCP = true;
}
else
{
ret = DataTransform.mrDataTransform(inst, inputMatrices, shuffleInst, otherInst, inst.getIv_resultIndices(), outputMatrices, inst.getIv_numReducers(), inst.getIv_replication());
}
break;
default:
throw new DMLRuntimeException("Invalid jobtype: " + inst.getJobType());
}
} // end of try block
catch (Exception e) {
throw new DMLRuntimeException( e );
}
if (ret.checkReturnStatus()) {
/*
* Check if any output is empty. If yes, create a dummy file. Needs
* to be done only in case of (1) CellOutputInfo and if not CP, or
* (2) BinaryBlockOutputInfo if not CP and output empty blocks disabled.
*/
try {
if( !execCP )
{
for (int i = 0; i < outputMatrices.length; i++) {
//get output meta data
MatrixFormatMetaData meta = (MatrixFormatMetaData)outputMatrices[i].getMetaData();
MatrixCharacteristics mc = meta.getMatrixCharacteristics();
OutputInfo outinfo = meta.getOutputInfo();
String fname = outputMatrices[i].getFileName();
if (MapReduceTool.isHDFSFileEmpty(fname))
{
//prepare output file
Path filepath = new Path(fname, "0-m-00000");
MatrixWriter writer = MatrixWriterFactory.createMatrixWriter(outinfo);
writer.writeEmptyMatrixToHDFS(filepath.toString(), mc.getRows(), mc.getCols(),
mc.getRowsPerBlock(), mc.getColsPerBlock());
}
outputMatrices[i].setHDFSFileExists(true);
if ( inst.getJobType() != JobType.CSV_WRITE && inst.getJobType() != JobType.TRANSFORM) {
// write out metadata file
// Currently, valueType information in not stored in MR instruction,
// since only DOUBLE matrices are supported ==> hard coded the value type information for now
MapReduceTool.writeMetaDataFile(fname + ".mtd", ValueType.DOUBLE, ((MatrixDimensionsMetaData)ret.getMetaData(i)).getMatrixCharacteristics(), outinfo);
}
}
}
return ret;
} catch (IOException e) {
throw new DMLRuntimeException(e);
}
}
// should not come here!
throw new DMLRuntimeException("Unexpected Job Type: " + inst.getJobType());
}
private static void checkEmptyInputs( MRJobInstruction inst, MatrixObject[] inputMatrices )
throws DMLRuntimeException
{
// Check if any of the input files are empty.. only for those job types
// for which empty inputs are NOT allowed
if (!inst.getJobType().areEmptyInputsAllowed()) {
for ( int i=0; i < inputMatrices.length; i++ ) {
try {
if (MapReduceTool.isHDFSFileEmpty(inputMatrices[i].getFileName())) {
throw new DMLRuntimeException( "Can not operate on an empty file: " + inputMatrices[i].getFileName());
}
} catch (IOException e) {
throw new DMLRuntimeException( "runtime error occurred -- " , e);
}
}
}
}
/**
* Computes the replacement string for a given variable name placeholder string
* (e.g., ##mVar2## or ##Var5##). The replacement is a HDFS filename for matrix
* variables, and is the actual value (stored in symbol table) for scalar variables.
*
* @param inst instruction
* @param varName variable name
* @param map local variable map
* @return string variable name
* @throws DMLRuntimeException if DMLRuntimeException occurs
*/
private static String getVarNameReplacement(String inst, String varName, LocalVariableMap map) throws DMLRuntimeException {
Data val = map.get(varName);
if ( val != null ) {
String replacement = null;
if (val.getDataType() == DataType.MATRIX) {
replacement = ((MatrixObject)val).getFileName();
}
if (val.getDataType() == DataType.SCALAR)
replacement = "" + ((ScalarObject) val).getStringValue();
return replacement;
}
else {
throw new DMLRuntimeException("Variable ("+varName+") in Instruction ("+inst+") is not found in the variablemap.");
}
}
/**
* Replaces ALL placeholder strings (such as ##mVar2## and ##Var5##) in a single instruction.
*
* @param inst string instruction
* @param map local variable map
* @return string instruction after replacement
* @throws DMLRuntimeException if DMLRuntimeException occurs
*/
private static String updateInstLabels(String inst, LocalVariableMap map) throws DMLRuntimeException {
if ( inst.contains(Lop.VARIABLE_NAME_PLACEHOLDER) ) {
int skip = Lop.VARIABLE_NAME_PLACEHOLDER.toString().length();
while ( inst.contains(Lop.VARIABLE_NAME_PLACEHOLDER) ) {
int startLoc = inst.indexOf(Lop.VARIABLE_NAME_PLACEHOLDER)+skip;
String varName = inst.substring(startLoc, inst.indexOf(Lop.VARIABLE_NAME_PLACEHOLDER, startLoc));
String replacement = getVarNameReplacement(inst, varName, map);
inst = inst.replaceAll(Lop.VARIABLE_NAME_PLACEHOLDER + varName + Lop.VARIABLE_NAME_PLACEHOLDER, replacement);
}
}
return inst;
}
/**
* Takes a delimited string of instructions, and replaces ALL placeholder labels
* (such as ##mVar2## and ##Var5##) in ALL instructions.
*
* @param instList instruction list as string
* @param labelValueMapping local variable map
* @return instruction list after replacement
* @throws DMLRuntimeException if DMLRuntimeException occurs
*/
public static String updateLabels (String instList, LocalVariableMap labelValueMapping) throws DMLRuntimeException {
if ( !instList.contains(Lop.VARIABLE_NAME_PLACEHOLDER) )
return instList;
StringBuilder updateInstList = new StringBuilder();
String[] ilist = instList.split(Lop.INSTRUCTION_DELIMITOR);
for ( int i=0; i < ilist.length; i++ ) {
if ( i > 0 )
updateInstList.append(Lop.INSTRUCTION_DELIMITOR);
updateInstList.append( updateInstLabels(ilist[i], labelValueMapping));
}
return updateInstList.toString();
}
private static long[] getNNZ( MatrixObject[] inputMatrices )
throws DMLRuntimeException
{
int len = inputMatrices.length;
long[] ret = new long[len];
for( int i=0; i<len; i++ )
{
MatrixObject mo = inputMatrices[i];
if( mo != null )
ret[i] = mo.getNnz();
else
ret[i] = -1;
}
return ret;
}
private static JobReturn executeInMemoryReblockOperations( MRJobInstruction inst, String shuffleInst, MatrixObject[] inputMatrices, MatrixObject[] outputMatrices )
throws DMLRuntimeException
{
MatrixCharacteristics[] mc = new MatrixCharacteristics[outputMatrices.length];
ReblockInstruction[] rblkSet = MRInstructionParser.parseReblockInstructions(shuffleInst);
byte[] results = inst.getIv_resultIndices();
for( ReblockInstruction rblk : rblkSet )
{
//CP Reblock through caching framework (no copy required: same data, next op copies)
MatrixBlock mb = inputMatrices[rblk.input].acquireRead();
for( int i=0; i<results.length; i++ )
if( rblk.output == results[i] )
{
outputMatrices[i].acquireModify( mb );
outputMatrices[i].release();
mc[i] = new MatrixCharacteristics(mb.getNumRows(),mb.getNumColumns(), rblk.brlen, rblk.bclen, mb.getNonZeros());
}
inputMatrices[rblk.input].release();
}
return new JobReturn( mc, inst.getOutputInfos(), true);
}
private static JobReturn executeInMemoryDataGenOperations( MRJobInstruction inst, String randInst, MatrixObject[] outputMatrices )
throws DMLRuntimeException
{
MatrixCharacteristics[] mc = new MatrixCharacteristics[outputMatrices.length];
DataGenMRInstruction[] dgSet = MRInstructionParser.parseDataGenInstructions(randInst);
byte[] results = inst.getIv_resultIndices();
for( DataGenMRInstruction ldgInst : dgSet )
{
if( ldgInst instanceof RandInstruction )
{
//CP Rand block operation
RandInstruction lrand = (RandInstruction)ldgInst;
RandomMatrixGenerator rgen = LibMatrixDatagen.createRandomMatrixGenerator(
lrand.getProbabilityDensityFunction(),
(int)lrand.getRows(), (int)lrand.getCols(),
lrand.getRowsInBlock(), lrand.getColsInBlock(),
lrand.getSparsity(), lrand.getMinValue(), lrand.getMaxValue(),
lrand.getPdfParams());
MatrixBlock mb = MatrixBlock.randOperations(rgen, lrand.getSeed());
for( int i=0; i<results.length; i++ )
if( lrand.output == results[i] )
{
outputMatrices[i].acquireModify( mb );
outputMatrices[i].release();
mc[i] = new MatrixCharacteristics(mb.getNumRows(),mb.getNumColumns(), lrand.getRowsInBlock(), lrand.getColsInBlock(), mb.getNonZeros());
}
}
else if( ldgInst instanceof SeqInstruction )
{
SeqInstruction lseq = (SeqInstruction) ldgInst;
MatrixBlock mb = MatrixBlock.seqOperations(lseq.fromValue, lseq.toValue, lseq.incrValue);
for( int i=0; i<results.length; i++ )
if( lseq.output == results[i] )
{
outputMatrices[i].acquireModify( mb );
outputMatrices[i].release();
mc[i] = new MatrixCharacteristics(mb.getNumRows(),mb.getNumColumns(), lseq.getRowsInBlock(), lseq.getColsInBlock(), mb.getNonZeros());
}
}
}
return new JobReturn( mc, inst.getOutputInfos(), true);
}
private static JobReturn executeInMemoryTransform( MRJobInstruction inst, MatrixObject[] inputMatrices, MatrixObject[] outputMatrices) throws IOException, DMLRuntimeException, IllegalArgumentException, JSONException {
return DataTransform.cpDataTransform(
inst.getIv_shuffleInstructions(),
inputMatrices,
outputMatrices);
}
}