Examples with SpillInfo org.apache.tez.runtime.library.common.writers.UnorderedPartitionedKVWriter.SpillInfo used on opensource projects

Example 1 with SpillInfo

use of org.apache.tez.runtime.library.common.writers.UnorderedPartitionedKVWriter.SpillInfo in project tez by apache.

the class TestUnorderedPartitionedKVWriter method baseTest.

private void baseTest(int numRecords, int numPartitions, Set<Integer> skippedPartitions, boolean shouldCompress, int maxSingleBufferSizeBytes, int bufferMergePercent, int availableMemory) throws IOException, InterruptedException {
    PartitionerForTest partitioner = new PartitionerForTest();
    ApplicationId appId = ApplicationId.newInstance(10000000, 1);
    TezCounters counters = new TezCounters();
    String uniqueId = UUID.randomUUID().toString();
    int dagId = 1;
    String auxiliaryService = defaultConf.get(TezConfiguration.TEZ_AM_SHUFFLE_AUXILIARY_SERVICE_ID, TezConfiguration.TEZ_AM_SHUFFLE_AUXILIARY_SERVICE_ID_DEFAULT);
    OutputContext outputContext = createMockOutputContext(counters, appId, uniqueId, auxiliaryService);
    Configuration conf = createConfiguration(outputContext, IntWritable.class, LongWritable.class, shouldCompress, maxSingleBufferSizeBytes);
    conf.setInt(TezRuntimeConfiguration.TEZ_RUNTIME_UNORDERED_PARTITIONED_KVWRITER_BUFFER_MERGE_PERCENT, bufferMergePercent);
    CompressionCodec codec = null;
    if (shouldCompress) {
        codec = new DefaultCodec();
        ((Configurable) codec).setConf(conf);
    }
    int numOutputs = numPartitions;
    int numRecordsWritten = 0;
    Map<Integer, Multimap<Integer, Long>> expectedValues = new HashMap<Integer, Multimap<Integer, Long>>();
    for (int i = 0; i < numOutputs; i++) {
        expectedValues.put(i, LinkedListMultimap.<Integer, Long>create());
    }
    UnorderedPartitionedKVWriter kvWriter = new UnorderedPartitionedKVWriterForTest(outputContext, conf, numOutputs, availableMemory);
    int sizePerBuffer = kvWriter.sizePerBuffer;
    // IntW + LongW
    int sizePerRecord = 4 + 8;
    // Record + META_OVERHEAD
    int sizePerRecordWithOverhead = sizePerRecord + 12;
    IntWritable intWritable = new IntWritable();
    LongWritable longWritable = new LongWritable();
    BitSet partitionsWithData = new BitSet(numPartitions);
    for (int i = 0; i < numRecords; i++) {
        intWritable.set(i);
        longWritable.set(i);
        int partition = partitioner.getPartition(intWritable, longWritable, numOutputs);
        if (skippedPartitions != null && skippedPartitions.contains(partition)) {
            continue;
        }
        partitionsWithData.set(partition);
        expectedValues.get(partition).put(intWritable.get(), longWritable.get());
        kvWriter.write(intWritable, longWritable);
        numRecordsWritten++;
    }
    List<Event> events = kvWriter.close();
    if (numPartitions == 1) {
        assertEquals(true, kvWriter.skipBuffers);
    }
    int recordsPerBuffer = sizePerBuffer / sizePerRecordWithOverhead;
    int numExpectedSpills = numRecordsWritten / recordsPerBuffer / kvWriter.spillLimit;
    verify(outputContext, never()).reportFailure(any(TaskFailureType.class), any(Throwable.class), any(String.class));
    assertNull(kvWriter.currentBuffer);
    assertEquals(0, kvWriter.availableBuffers.size());
    // Verify the counters
    TezCounter outputRecordBytesCounter = counters.findCounter(TaskCounter.OUTPUT_BYTES);
    TezCounter outputRecordsCounter = counters.findCounter(TaskCounter.OUTPUT_RECORDS);
    TezCounter outputBytesWithOverheadCounter = counters.findCounter(TaskCounter.OUTPUT_BYTES_WITH_OVERHEAD);
    TezCounter fileOutputBytesCounter = counters.findCounter(TaskCounter.OUTPUT_BYTES_PHYSICAL);
    TezCounter spilledRecordsCounter = counters.findCounter(TaskCounter.SPILLED_RECORDS);
    TezCounter additionalSpillBytesWritternCounter = counters.findCounter(TaskCounter.ADDITIONAL_SPILLS_BYTES_WRITTEN);
    TezCounter additionalSpillBytesReadCounter = counters.findCounter(TaskCounter.ADDITIONAL_SPILLS_BYTES_READ);
    TezCounter numAdditionalSpillsCounter = counters.findCounter(TaskCounter.ADDITIONAL_SPILL_COUNT);
    assertEquals(numRecordsWritten * sizePerRecord, outputRecordBytesCounter.getValue());
    if (numPartitions > 1) {
        assertEquals(numRecordsWritten * sizePerRecordWithOverhead, outputBytesWithOverheadCounter.getValue());
    }
    assertEquals(numRecordsWritten, outputRecordsCounter.getValue());
    long fileOutputBytes = fileOutputBytesCounter.getValue();
    if (numRecordsWritten > 0) {
        assertTrue(fileOutputBytes > 0);
        if (!shouldCompress) {
            assertTrue(fileOutputBytes > outputRecordBytesCounter.getValue());
        }
    } else {
        assertEquals(0, fileOutputBytes);
    }
    assertEquals(recordsPerBuffer * numExpectedSpills, spilledRecordsCounter.getValue());
    long additionalSpillBytesWritten = additionalSpillBytesWritternCounter.getValue();
    long additionalSpillBytesRead = additionalSpillBytesReadCounter.getValue();
    if (numExpectedSpills == 0) {
        assertEquals(0, additionalSpillBytesWritten);
        assertEquals(0, additionalSpillBytesRead);
    } else {
        assertTrue(additionalSpillBytesWritten > 0);
        assertTrue(additionalSpillBytesRead > 0);
        if (!shouldCompress) {
            assertTrue(additionalSpillBytesWritten > (recordsPerBuffer * numExpectedSpills * sizePerRecord));
            assertTrue(additionalSpillBytesRead > (recordsPerBuffer * numExpectedSpills * sizePerRecord));
        }
    }
    assertEquals(additionalSpillBytesWritten, additionalSpillBytesRead);
    // due to multiple threads, buffers could be merged in chunks in scheduleSpill.
    assertTrue(numExpectedSpills >= numAdditionalSpillsCounter.getValue());
    BitSet emptyPartitionBits = null;
    // Verify the events returned
    assertEquals(2, events.size());
    assertTrue(events.get(0) instanceof VertexManagerEvent);
    VertexManagerEvent vme = (VertexManagerEvent) events.get(0);
    verifyPartitionStats(vme, partitionsWithData);
    assertTrue(events.get(1) instanceof CompositeDataMovementEvent);
    CompositeDataMovementEvent cdme = (CompositeDataMovementEvent) events.get(1);
    assertEquals(0, cdme.getSourceIndexStart());
    assertEquals(numOutputs, cdme.getCount());
    DataMovementEventPayloadProto eventProto = DataMovementEventPayloadProto.parseFrom(ByteString.copyFrom(cdme.getUserPayload()));
    if (skippedPartitions == null && numRecordsWritten > 0) {
        assertFalse(eventProto.hasEmptyPartitions());
        emptyPartitionBits = new BitSet(numPartitions);
    } else {
        assertTrue(eventProto.hasEmptyPartitions());
        byte[] emptyPartitions = TezCommonUtils.decompressByteStringToByteArray(eventProto.getEmptyPartitions());
        emptyPartitionBits = TezUtilsInternal.fromByteArray(emptyPartitions);
        if (numRecordsWritten == 0) {
            assertEquals(numPartitions, emptyPartitionBits.cardinality());
        } else {
            for (Integer e : skippedPartitions) {
                assertTrue(emptyPartitionBits.get(e));
            }
            assertEquals(skippedPartitions.size(), emptyPartitionBits.cardinality());
        }
    }
    if (emptyPartitionBits.cardinality() != numPartitions) {
        assertEquals(HOST_STRING, eventProto.getHost());
        assertEquals(SHUFFLE_PORT, eventProto.getPort());
        assertEquals(uniqueId, eventProto.getPathComponent());
    } else {
        assertFalse(eventProto.hasHost());
        assertFalse(eventProto.hasPort());
        assertFalse(eventProto.hasPathComponent());
    }
    // Verify the actual data
    TezTaskOutput taskOutput = new TezTaskOutputFiles(conf, uniqueId, dagId);
    Path outputFilePath = kvWriter.finalOutPath;
    Path spillFilePath = kvWriter.finalIndexPath;
    if (numRecordsWritten <= 0) {
        return;
    }
    assertTrue(localFs.exists(outputFilePath));
    assertTrue(localFs.exists(spillFilePath));
    // verify no intermediate spill files have been left around
    synchronized (kvWriter.spillInfoList) {
        for (SpillInfo spill : kvWriter.spillInfoList) {
            assertFalse("lingering intermediate spill file " + spill.outPath, localFs.exists(spill.outPath));
        }
    }
    // Special case for 0 records.
    TezSpillRecord spillRecord = new TezSpillRecord(spillFilePath, conf);
    DataInputBuffer keyBuffer = new DataInputBuffer();
    DataInputBuffer valBuffer = new DataInputBuffer();
    IntWritable keyDeser = new IntWritable();
    LongWritable valDeser = new LongWritable();
    for (int i = 0; i < numOutputs; i++) {
        TezIndexRecord indexRecord = spillRecord.getIndex(i);
        if (skippedPartitions != null && skippedPartitions.contains(i)) {
            assertFalse("The Index Record for partition " + i + " should not have any data", indexRecord.hasData());
            continue;
        }
        FSDataInputStream inStream = FileSystem.getLocal(conf).open(outputFilePath);
        inStream.seek(indexRecord.getStartOffset());
        IFile.Reader reader = new IFile.Reader(inStream, indexRecord.getPartLength(), codec, null, null, false, 0, -1);
        while (reader.nextRawKey(keyBuffer)) {
            reader.nextRawValue(valBuffer);
            keyDeser.readFields(keyBuffer);
            valDeser.readFields(valBuffer);
            int partition = partitioner.getPartition(keyDeser, valDeser, numOutputs);
            assertTrue(expectedValues.get(partition).remove(keyDeser.get(), valDeser.get()));
        }
        inStream.close();
    }
    for (int i = 0; i < numOutputs; i++) {
        assertEquals(0, expectedValues.get(i).size());
        expectedValues.remove(i);
    }
    assertEquals(0, expectedValues.size());
    verify(outputContext, atLeast(1)).notifyProgress();
}