Satish Subhashrao Saley created PIG-5359: --------------------------------------------
Summary: Reduce time spent in split serialization Key: PIG-5359 URL: https://issues.apache.org/jira/browse/PIG-5359 Project: Pig Issue Type: Improvement Reporter: Satish Subhashrao Saley Assignee: Satish Subhashrao Saley 1. Unnecessary serialization of splits in Tez. In LoaderProcessor, pig calls [https://github.com/apache/pig/blob/trunk/src/org/apache/pig/backend/hadoop/executionengine/tez/plan/optimizer/LoaderProcessor.java#L172] {code:java} tezOp.getLoaderInfo().setInputSplitInfo(MRInputHelpers.generateInputSplitsToMem(conf, false, 0)); {code} It ends up serializing the splits, just to print log. [https://github.com/apache/tez/blob/master/tez-mapreduce/src/main/java/org/apache/tez/mapreduce/hadoop/MRInputHelpers.java#L317] {code:java} public static InputSplitInfoMem generateInputSplitsToMem(Configuration conf, boolean groupSplits, boolean sortSplits, int targetTasks) throws IOException, ClassNotFoundException, InterruptedException { .... .... LOG.info("NumSplits: " + splitInfoMem.getNumTasks() + ", SerializedSize: " + splitInfoMem.getSplitsProto().getSerializedSize()); return splitInfoMem; {code} [https://github.com/apache/tez/blob/master/tez-mapreduce/src/main/java/org/apache/tez/mapreduce/hadoop/InputSplitInfoMem.java#L106] {code:java} public MRSplitsProto getSplitsProto() { if (isNewSplit) { try { return createSplitsProto(newFormatSplits, new SerializationFactory(conf)); {code} [https://github.com/apache/tez/blob/master/tez-mapreduce/src/main/java/org/apache/tez/mapreduce/hadoop/InputSplitInfoMem.java#L152-L170] {code:java} private static MRSplitsProto createSplitsProto( org.apache.hadoop.mapreduce.InputSplit[] newSplits, SerializationFactory serializationFactory) throws IOException, InterruptedException { MRSplitsProto.Builder splitsBuilder = MRSplitsProto.newBuilder(); for (org.apache.hadoop.mapreduce.InputSplit newSplit : newSplits) { splitsBuilder.addSplits(MRInputHelpers.createSplitProto(newSplit, serializationFactory)); } return splitsBuilder.build(); } {code} [https://github.com/apache/tez/blob/master/tez-mapreduce/src/main/java/org/apache/tez/mapreduce/hadoop/MRInputHelpers.java#L221-L259] 2. In TezDagBuilder, if splitsSerializedSize > spillThreshold, then the InputSplits serialized in MRSplitsProto are not used by Pig and it serializes again directly to disk via JobSplitWriter.createSplitFiles. So the InputSplit serialization logic is called again which is wasteful and expensive in cases like HCat. [https://github.com/apache/pig/blob/trunk/src/org/apache/pig/backend/hadoop/executionengine/tez/TezDagBuilder.java#L946-L947] {code:java} MRSplitsProto splitsProto = inputSplitInfo.getSplitsProto(); int splitsSerializedSize = splitsProto.getSerializedSize(); {code} The getSplitsProto, creates MRSplitsProto which consists of list of MRSplitProto. MRSplitProto has serialized bytes of each InputFormat. If splitsSerializedSize > spillThreshold, pig writes the splits to disk via {code:java} if(splitsSerializedSize > spillThreshold) { inputPayLoad.setBoolean( org.apache.tez.mapreduce.hadoop.MRJobConfig.MR_TEZ_SPLITS_VIA_EVENTS, false); // Write splits to disk Path inputSplitsDir = FileLocalizer.getTemporaryPath(pc); log.info("Writing input splits to " + inputSplitsDir + " for vertex " + vertex.getName() + " as the serialized size in memory is " + splitsSerializedSize + ". Configured " + PigConfiguration.PIG_TEZ_INPUT_SPLITS_MEM_THRESHOLD + " is " + spillThreshold); inputSplitInfo = MRToTezHelper.writeInputSplitInfoToDisk( (InputSplitInfoMem)inputSplitInfo, inputSplitsDir, payloadConf, fs); {code} [https://github.com/apache/pig/blob/trunk/src/org/apache/pig/backend/hadoop/executionengine/tez/TezDagBuilder.java#L960] [https://github.com/apache/pig/blob/trunk/src/org/apache/pig/backend/hadoop/executionengine/tez/util/MRToTezHelper.java#L302-L314] Solution: 1. Do not serialize the split in LoaderProcessor.java 2. In TezDagBuilder.java, serialize each input split and keep adding its size and if it exceeds spillThreshold, then write the splits to disk reusing the serialized buffers for each split. Thank you [~rohini] for identifying the issue. -- This message was sent by Atlassian JIRA (v7.6.3#76005)