use of org.apache.sysml.conf.DMLConfig in project incubator-systemml by apache.
the class ProgramRecompiler method generatePartitialRuntimeProgram.
public static ArrayList<ProgramBlock> generatePartitialRuntimeProgram(Program rtprog, ArrayList<StatementBlock> sbs) throws LopsException, DMLRuntimeException, IOException, HopsException {
ArrayList<ProgramBlock> ret = new ArrayList<ProgramBlock>();
DMLConfig config = ConfigurationManager.getDMLConfig();
//construct lops from hops if not existing
DMLTranslator dmlt = new DMLTranslator(sbs.get(0).getDMLProg());
for (StatementBlock sb : sbs) {
dmlt.constructLops(sb);
}
//construct runtime program from lops
for (StatementBlock sb : sbs) {
DMLProgram prog = sb.getDMLProg();
ret.add(prog.createRuntimeProgramBlock(rtprog, sb, config));
}
return ret;
}
use of org.apache.sysml.conf.DMLConfig in project incubator-systemml by apache.
the class LocalFileUtils method createWorkingDirectoryWithUUID.
public static String createWorkingDirectoryWithUUID(String uuid) throws DMLRuntimeException {
//create local tmp dir if not existing
String dirRoot = null;
DMLConfig conf = ConfigurationManager.getDMLConfig();
if (conf != null)
dirRoot = conf.getTextValue(DMLConfig.LOCAL_TMP_DIR);
else
dirRoot = DMLConfig.getDefaultTextValue(DMLConfig.LOCAL_TMP_DIR);
//create shared staging dir if not existing
if (!LocalFileUtils.createLocalFileIfNotExist(dirRoot, DMLConfig.DEFAULT_SHARED_DIR_PERMISSION)) {
throw new DMLRuntimeException("Failed to create non-existing local working directory: " + dirRoot);
}
//create process specific sub tmp dir
StringBuilder sb = new StringBuilder();
sb.append(dirRoot);
sb.append(Lop.FILE_SEPARATOR);
sb.append(Lop.PROCESS_PREFIX);
sb.append(uuid);
sb.append(Lop.FILE_SEPARATOR);
_workingDir = sb.toString();
//create process-specific staging dir if not existing
if (!LocalFileUtils.createLocalFileIfNotExist(_workingDir)) {
throw new DMLRuntimeException("Failed to create local working directory: " + _workingDir);
}
return _workingDir;
}
use of org.apache.sysml.conf.DMLConfig in project incubator-systemml by apache.
the class DataPartitionerRemoteMR method partitionMatrix.
@Override
protected void partitionMatrix(MatrixObject in, String fnameNew, InputInfo ii, OutputInfo oi, long rlen, long clen, int brlen, int bclen) throws DMLRuntimeException {
String jobname = "ParFor-DPMR";
long t0 = DMLScript.STATISTICS ? System.nanoTime() : 0;
JobConf job;
job = new JobConf(DataPartitionerRemoteMR.class);
if (//use in parfor
_pfid >= 0)
job.setJobName(jobname + _pfid);
else
//use for partition instruction
job.setJobName("Partition-MR");
//maintain dml script counters
Statistics.incrementNoOfCompiledMRJobs();
try {
//force writing to disk (typically not required since partitioning only applied if dataset exceeds CP size)
//written to disk iff dirty
in.exportData();
Path path = new Path(in.getFileName());
/////
//configure the MR job
MRJobConfiguration.setPartitioningInfo(job, rlen, clen, brlen, bclen, ii, oi, _format, _n, fnameNew, _keepIndexes);
//set mappers, reducers, combiners
job.setMapperClass(DataPartitionerRemoteMapper.class);
job.setReducerClass(DataPartitionerRemoteReducer.class);
if (oi == OutputInfo.TextCellOutputInfo) {
//binary cell intermediates for reduced IO
job.setMapOutputKeyClass(LongWritable.class);
job.setMapOutputValueClass(PairWritableCell.class);
} else if (oi == OutputInfo.BinaryCellOutputInfo) {
job.setMapOutputKeyClass(LongWritable.class);
job.setMapOutputValueClass(PairWritableCell.class);
} else if (oi == OutputInfo.BinaryBlockOutputInfo) {
job.setMapOutputKeyClass(LongWritable.class);
job.setMapOutputValueClass(PairWritableBlock.class);
//check Alignment
if ((_format == PDataPartitionFormat.ROW_BLOCK_WISE_N && rlen > _n && _n % brlen != 0) || (_format == PDataPartitionFormat.COLUMN_BLOCK_WISE_N && clen > _n && _n % bclen != 0)) {
throw new DMLRuntimeException("Data partitioning format " + _format + " requires aligned blocks.");
}
}
//set input format
job.setInputFormat(ii.inputFormatClass);
//set the input path and output path
FileInputFormat.setInputPaths(job, path);
//set output path
MapReduceTool.deleteFileIfExistOnHDFS(fnameNew);
//FileOutputFormat.setOutputPath(job, pathNew);
job.setOutputFormat(NullOutputFormat.class);
//////
//set optimization parameters
//set the number of mappers and reducers
//job.setNumMapTasks( _numMappers ); //use default num mappers
long reducerGroups = -1;
switch(_format) {
case ROW_WISE:
reducerGroups = rlen;
break;
case COLUMN_WISE:
reducerGroups = clen;
break;
case ROW_BLOCK_WISE:
reducerGroups = (rlen / brlen) + ((rlen % brlen == 0) ? 0 : 1);
break;
case COLUMN_BLOCK_WISE:
reducerGroups = (clen / bclen) + ((clen % bclen == 0) ? 0 : 1);
break;
case ROW_BLOCK_WISE_N:
reducerGroups = (rlen / _n) + ((rlen % _n == 0) ? 0 : 1);
break;
case COLUMN_BLOCK_WISE_N:
reducerGroups = (clen / _n) + ((clen % _n == 0) ? 0 : 1);
break;
default:
}
job.setNumReduceTasks((int) Math.min(_numReducers, reducerGroups));
//disable automatic tasks timeouts and speculative task exec
job.setInt(MRConfigurationNames.MR_TASK_TIMEOUT, 0);
job.setMapSpeculativeExecution(false);
//set up preferred custom serialization framework for binary block format
if (MRJobConfiguration.USE_BINARYBLOCK_SERIALIZATION)
MRJobConfiguration.addBinaryBlockSerializationFramework(job);
//enables the reuse of JVMs (multiple tasks per MR task)
if (_jvmReuse)
//unlimited
job.setNumTasksToExecutePerJvm(-1);
//enables compression - not conclusive for different codecs (empirically good compression ratio, but significantly slower)
//job.set(MRConfigurationNames.MR_MAP_OUTPUT_COMPRESS, "true");
//job.set(MRConfigurationNames.MR_MAP_OUTPUT_COMPRESS_CODEC, "org.apache.hadoop.io.compress.GzipCodec");
//set the replication factor for the results
job.setInt(MRConfigurationNames.DFS_REPLICATION, _replication);
//set up map/reduce memory configurations (if in AM context)
DMLConfig config = ConfigurationManager.getDMLConfig();
DMLAppMasterUtils.setupMRJobRemoteMaxMemory(job, config);
//set up custom map/reduce configurations
MRJobConfiguration.setupCustomMRConfigurations(job, config);
//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(MRConfigurationNames.MR_MAP_MAXATTEMPTS, _max_retry);
//set unique working dir
MRJobConfiguration.setUniqueWorkingDir(job);
/////
// execute the MR job
JobClient.runJob(job);
//maintain dml script counters
Statistics.incrementNoOfExecutedMRJobs();
} catch (Exception ex) {
throw new DMLRuntimeException(ex);
}
if (DMLScript.STATISTICS && _pfid >= 0) {
//only for parfor
long t1 = System.nanoTime();
Statistics.maintainCPHeavyHitters("MR-Job_" + jobname, t1 - t0);
}
}
use of org.apache.sysml.conf.DMLConfig in project incubator-systemml by apache.
the class ReblockMR method runJob.
public static JobReturn runJob(MRJobInstruction inst, String[] inputs, InputInfo[] inputInfos, long[] rlens, long[] clens, int[] brlens, int[] bclens, long[] nnz, String instructionsInMapper, String reblockInstructions, String otherInstructionsInReducer, int numReducers, int replication, boolean jvmReuse, byte[] resultIndexes, String[] outputs, OutputInfo[] outputInfos) throws Exception {
JobConf job = new JobConf(ReblockMR.class);
job.setJobName("Reblock-MR");
byte[] realIndexes = new byte[inputs.length];
for (byte b = 0; b < realIndexes.length; b++) realIndexes[b] = b;
//set up the input files and their format information
//(internally used input converters: text2bc for text, identity for binary inputs)
MRJobConfiguration.setUpMultipleInputsReblock(job, realIndexes, inputs, inputInfos, brlens, bclens);
//set up the dimensions of input matrices
MRJobConfiguration.setMatricesDimensions(job, realIndexes, rlens, clens, nnz);
//set up the block size
MRJobConfiguration.setBlocksSizes(job, realIndexes, brlens, bclens);
//set up unary instructions that will perform in the mapper
MRJobConfiguration.setInstructionsInMapper(job, instructionsInMapper);
//set up the aggregate instructions that will happen in the combiner and reducer
MRJobConfiguration.setReblockInstructions(job, reblockInstructions);
//set up the instructions that will happen in the reducer, after the aggregation instrucions
MRJobConfiguration.setInstructionsInReducer(job, otherInstructionsInReducer);
//set up the replication factor for the results
job.setInt(MRConfigurationNames.DFS_REPLICATION, replication);
//disable automatic tasks timeouts and speculative task exec
job.setInt(MRConfigurationNames.MR_TASK_TIMEOUT, 0);
job.setMapSpeculativeExecution(false);
//set up preferred custom serialization framework for binary block format
if (MRJobConfiguration.USE_BINARYBLOCK_SERIALIZATION)
MRJobConfiguration.addBinaryBlockSerializationFramework(job);
//set up custom map/reduce configurations
DMLConfig config = ConfigurationManager.getDMLConfig();
MRJobConfiguration.setupCustomMRConfigurations(job, config);
//enable jvm reuse (based on SystemML configuration)
if (jvmReuse)
job.setNumTasksToExecutePerJvm(-1);
//set up what matrices are needed to pass from the mapper to reducer
HashSet<Byte> mapoutputIndexes = MRJobConfiguration.setUpOutputIndexesForMapper(job, realIndexes, instructionsInMapper, reblockInstructions, null, otherInstructionsInReducer, resultIndexes);
MatrixChar_N_ReducerGroups ret = MRJobConfiguration.computeMatrixCharacteristics(job, realIndexes, instructionsInMapper, reblockInstructions, null, null, otherInstructionsInReducer, resultIndexes, mapoutputIndexes, false);
MatrixCharacteristics[] stats = ret.stats;
//set up the number of reducers (according to output size)
int numRed = determineNumReducers(rlens, clens, nnz, config.getIntValue(DMLConfig.NUM_REDUCERS), ret.numReducerGroups);
job.setNumReduceTasks(numRed);
// Print the complete instruction
if (LOG.isTraceEnabled())
inst.printCompleteMRJobInstruction(stats);
// Update resultDimsUnknown based on computed "stats"
byte[] resultDimsUnknown = new byte[resultIndexes.length];
for (int i = 0; i < resultIndexes.length; i++) {
if (stats[i].getRows() == -1 || stats[i].getCols() == -1) {
resultDimsUnknown[i] = (byte) 1;
} else {
resultDimsUnknown[i] = (byte) 0;
}
}
//set up the multiple output files, and their format information
MRJobConfiguration.setUpMultipleOutputs(job, resultIndexes, resultDimsUnknown, outputs, outputInfos, true, true);
// configure mapper and the mapper output key value pairs
job.setMapperClass(ReblockMapper.class);
//represent key offsets for block
job.setMapOutputKeyClass(MatrixIndexes.class);
//binary cell/block
job.setMapOutputValueClass(TaggedAdaptivePartialBlock.class);
//configure reducer
job.setReducerClass(ReblockReducer.class);
// By default, the job executes in "cluster" mode.
// Determine if we can optimize and run it in "local" mode.
// at this point, both reblock_binary and reblock_text are similar
MatrixCharacteristics[] inputStats = new MatrixCharacteristics[inputs.length];
for (int i = 0; i < inputs.length; i++) {
inputStats[i] = new MatrixCharacteristics(rlens[i], clens[i], brlens[i], bclens[i]);
}
//set unique working dir
MRJobConfiguration.setUniqueWorkingDir(job);
RunningJob runjob = JobClient.runJob(job);
/* Process different counters */
Group group = runjob.getCounters().getGroup(MRJobConfiguration.NUM_NONZERO_CELLS);
for (int i = 0; i < resultIndexes.length; i++) {
// number of non-zeros
stats[i].setNonZeros(group.getCounter(Integer.toString(i)));
// System.out.println("result #"+resultIndexes[i]+" ===>\n"+stats[i]);
}
return new JobReturn(stats, outputInfos, runjob.isSuccessful());
}
use of org.apache.sysml.conf.DMLConfig in project incubator-systemml by apache.
the class SortMR method runJob.
@SuppressWarnings({ "unchecked", "rawtypes" })
public static JobReturn runJob(MRJobInstruction inst, String input, InputInfo inputInfo, long rlen, long clen, int brlen, int bclen, String combineInst, String sortInst, int numReducers, int replication, String output, OutputInfo outputInfo, boolean valueIsWeight) throws Exception {
boolean sortIndexes = getSortInstructionType(sortInst) == SortKeys.OperationTypes.Indexes;
String tmpOutput = sortIndexes ? MRJobConfiguration.constructTempOutputFilename() : output;
JobConf job = new JobConf(SortMR.class);
job.setJobName("SortMR");
//setup partition file
String pfname = MRJobConfiguration.setUpSortPartitionFilename(job);
Path partitionFile = new Path(pfname);
URI partitionUri = new URI(partitionFile.toString());
//setup input/output paths
Path inputDir = new Path(input);
inputDir = inputDir.makeQualified(inputDir.getFileSystem(job));
SamplingSortMRInputFormat.setInputPaths(job, inputDir);
Path outpath = new Path(tmpOutput);
FileOutputFormat.setOutputPath(job, outpath);
MapReduceTool.deleteFileIfExistOnHDFS(outpath, job);
//set number of reducers (1 if local mode)
if (!InfrastructureAnalyzer.isLocalMode(job)) {
MRJobConfiguration.setNumReducers(job, numReducers, numReducers);
//on cp-side qpick instructions for quantile/iqm/median (~128MB)
if (!(getSortInstructionType(sortInst) == SortKeys.OperationTypes.Indexes))
job.setNumReduceTasks((int) Math.max(job.getNumReduceTasks(), rlen / 10000000));
} else
//in case of local mode
job.setNumReduceTasks(1);
//setup input/output format
job.setInputFormat(SamplingSortMRInputFormat.class);
SamplingSortMRInputFormat.setTargetKeyValueClasses(job, (Class<? extends WritableComparable>) outputInfo.outputKeyClass, outputInfo.outputValueClass);
//setup instructions and meta information
if (combineInst != null && !combineInst.trim().isEmpty())
job.set(COMBINE_INSTRUCTION, combineInst);
job.set(SORT_INSTRUCTION, sortInst);
job.setBoolean(VALUE_IS_WEIGHT, valueIsWeight);
boolean desc = getSortInstructionDescending(sortInst);
job.setBoolean(SORT_DECREASING, desc);
MRJobConfiguration.setBlockSize(job, (byte) 0, brlen, bclen);
MRJobConfiguration.setInputInfo(job, (byte) 0, inputInfo, brlen, bclen, ConvertTarget.CELL);
int partitionWith0 = SamplingSortMRInputFormat.writePartitionFile(job, partitionFile);
//setup mapper/reducer/partitioner/output classes
if (getSortInstructionType(sortInst) == SortKeys.OperationTypes.Indexes) {
MRJobConfiguration.setInputInfo(job, (byte) 0, inputInfo, brlen, bclen, ConvertTarget.CELL);
job.setOutputFormat(OutputInfo.BinaryBlockOutputInfo.outputFormatClass);
job.setMapperClass(IndexSortMapper.class);
job.setReducerClass(IndexSortReducer.class);
job.setMapOutputKeyClass(!desc ? IndexSortComparable.class : IndexSortComparableDesc.class);
job.setMapOutputValueClass(LongWritable.class);
job.setOutputKeyClass(MatrixIndexes.class);
job.setOutputValueClass(MatrixBlock.class);
} else {
//default case: SORT w/wo weights
MRJobConfiguration.setInputInfo(job, (byte) 0, inputInfo, brlen, bclen, ConvertTarget.CELL);
job.setOutputFormat(CompactOutputFormat.class);
job.setMapperClass(ValueSortMapper.class);
job.setReducerClass(ValueSortReducer.class);
//double
job.setOutputKeyClass(outputInfo.outputKeyClass);
//int
job.setOutputValueClass(outputInfo.outputValueClass);
}
job.setPartitionerClass(TotalOrderPartitioner.class);
//setup distributed cache
DistributedCache.addCacheFile(partitionUri, job);
DistributedCache.createSymlink(job);
//setup replication factor
job.setInt(MRConfigurationNames.DFS_REPLICATION, replication);
//set up custom map/reduce configurations
DMLConfig config = ConfigurationManager.getDMLConfig();
MRJobConfiguration.setupCustomMRConfigurations(job, config);
MatrixCharacteristics[] s = new MatrixCharacteristics[1];
s[0] = new MatrixCharacteristics(rlen, clen, brlen, bclen);
// Print the complete instruction
if (LOG.isTraceEnabled())
inst.printCompleteMRJobInstruction(s);
//set unique working dir
MRJobConfiguration.setUniqueWorkingDir(job);
//run mr job
RunningJob runjob = JobClient.runJob(job);
Group group = runjob.getCounters().getGroup(NUM_VALUES_PREFIX);
numReducers = job.getNumReduceTasks();
//process final meta data
long[] counts = new long[numReducers];
long total = 0;
for (int i = 0; i < numReducers; i++) {
counts[i] = group.getCounter(Integer.toString(i));
total += counts[i];
}
//add missing 0s back to the results
long missing0s = 0;
if (total < rlen * clen) {
if (partitionWith0 < 0)
throw new RuntimeException("no partition contains 0, which is wrong!");
missing0s = rlen * clen - total;
counts[partitionWith0] += missing0s;
} else
partitionWith0 = -1;
if (sortIndexes) {
//run builtin job for shifting partially sorted blocks according to global offsets
//we do this in this custom form since it would not fit into the current structure
//of systemml to output two intermediates (partially sorted data, offsets) out of a
//single SortKeys lop
boolean success = runjob.isSuccessful();
if (success) {
success = runStitchupJob(tmpOutput, rlen, clen, brlen, bclen, counts, numReducers, replication, output);
}
MapReduceTool.deleteFileIfExistOnHDFS(tmpOutput);
MapReduceTool.deleteFileIfExistOnHDFS(pfname);
return new JobReturn(s[0], OutputInfo.BinaryBlockOutputInfo, success);
} else {
MapReduceTool.deleteFileIfExistOnHDFS(pfname);
return new JobReturn(s[0], counts, partitionWith0, missing0s, runjob.isSuccessful());
}
}
Aggregations