Examples with TaskPartitioner edu.iu.dsc.tws.api.compute.TaskPartitioner used on opensource projects

Example 1 with TaskPartitioner

use of edu.iu.dsc.tws.api.compute.TaskPartitioner in project twister2 by DSC-SPIDAL.

the class TeraSort method execute.

@Override
public void execute(WorkerEnvironment workerEnv) {
    int workerID = workerEnv.getWorkerId();
    ComputeEnvironment cEnv = ComputeEnvironment.init(workerEnv);
    Config config = workerEnv.getConfig();
    resultsRecorder = new BenchmarkResultsRecorder(config, workerID == 0);
    Timing.setDefaultTimingUnit(TimingUnit.MILLI_SECONDS);
    final String filePath = config.getStringValue(ARG_INPUT_FILE, null);
    final int keySize = config.getIntegerValue(ARG_KEY_SIZE, 10);
    final int valueSize = config.getIntegerValue(ARG_VALUE_SIZE, 90);
    // Sampling Graph : if file based only
    TaskPartitioner taskPartitioner;
    if (filePath != null) {
        ComputeGraphBuilder samplingGraph = ComputeGraphBuilder.newBuilder(config);
        samplingGraph.setMode(OperationMode.BATCH);
        Sampler samplerTask = new Sampler();
        samplingGraph.addSource(TASK_SAMPLER, samplerTask, config.getIntegerValue(ARG_TASKS_SOURCES, 4));
        SamplerReduce samplerReduce = new SamplerReduce();
        samplingGraph.addCompute(TASK_SAMPLER_REDUCE, samplerReduce, config.getIntegerValue(ARG_RESOURCE_INSTANCES, 4)).allreduce(TASK_SAMPLER).viaEdge(EDGE).withReductionFunction(byte[].class, (minMax1, minMax2) -> {
            byte[] min1 = Arrays.copyOfRange(minMax1, 0, keySize);
            byte[] max1 = Arrays.copyOfRange(minMax1, keySize, minMax1.length);
            byte[] min2 = Arrays.copyOfRange(minMax2, 0, keySize);
            byte[] max2 = Arrays.copyOfRange(minMax2, keySize, minMax2.length);
            byte[] newMinMax = new byte[keySize * 2];
            byte[] min = min1;
            byte[] max = max1;
            if (ByteArrayComparator.getInstance().compare(min1, min2) > 0) {
                min = min2;
            }
            if (ByteArrayComparator.getInstance().compare(max1, max2) < 0) {
                max = max2;
            }
            System.arraycopy(min, 0, newMinMax, 0, keySize);
            System.arraycopy(max, 0, newMinMax, keySize, keySize);
            return newMinMax;
        });
        ComputeGraph sampleGraphBuild = samplingGraph.build();
        ExecutionPlan sampleTaskPlan = cEnv.getTaskExecutor().plan(sampleGraphBuild);
        cEnv.getTaskExecutor().execute(sampleGraphBuild, sampleTaskPlan);
        DataObject<byte[]> output = cEnv.getTaskExecutor().getOutput("sample-reduce");
        LOG.info("Sample output received");
        taskPartitioner = new TaskPartitionerForSampledData(output.getPartitions()[0].getConsumer().next(), keySize);
    } else {
        taskPartitioner = new TaskPartitionerForRandom();
    }
    // Sort Graph
    ComputeGraphBuilder teraSortTaskGraph = ComputeGraphBuilder.newBuilder(config);
    teraSortTaskGraph.setMode(OperationMode.BATCH);
    BaseSource dataSource;
    if (filePath == null) {
        dataSource = new RandomDataSource();
    } else {
        dataSource = new FileDataSource();
    }
    teraSortTaskGraph.addSource(TASK_SOURCE, dataSource, config.getIntegerValue(ARG_TASKS_SOURCES, 4));
    Receiver receiver = new Receiver();
    KeyedGatherConfig keyedGatherConfig = teraSortTaskGraph.addCompute(TASK_RECV, receiver, config.getIntegerValue(ARG_TASKS_SINKS, 4)).keyedGather(TASK_SOURCE).viaEdge(EDGE).withDataType(MessageTypes.BYTE_ARRAY).withKeyType(MessageTypes.BYTE_ARRAY).withTaskPartitioner(taskPartitioner).useDisk(true).sortBatchByKey(ByteArrayComparator.getInstance()).groupBatchByKey(false);
    if (config.getBooleanValue(ARG_FIXED_SCHEMA, false)) {
        LOG.info("Using fixed schema feature with message size : " + (keySize + valueSize) + " and key size : " + keySize);
        keyedGatherConfig.withMessageSchema(MessageSchema.ofSize(keySize + valueSize, keySize));
    }
    ComputeGraph computeGraph = teraSortTaskGraph.build();
    ExecutionPlan executionPlan = cEnv.getTaskExecutor().plan(computeGraph);
    cEnv.getTaskExecutor().execute(computeGraph, executionPlan);
    cEnv.close();
    LOG.info("Finished Sorting...");
}

Example 2 with TaskPartitioner

use of edu.iu.dsc.tws.api.compute.TaskPartitioner in project twister2 by DSC-SPIDAL.

the class BTJoinExample method buildTaskGraph.

@Override
public ComputeGraphBuilder buildTaskGraph() {
    List<Integer> taskStages = jobParameters.getTaskStages();
    int sourceParallelism = taskStages.get(0);
    int sinkParallelism = taskStages.get(1);
    MessageType keyType = MessageTypes.INTEGER;
    MessageType dataType = MessageTypes.INTEGER_ARRAY;
    BaseSource source1 = new JoinSource(JoinRelation.LEFT);
    BaseSource source2 = new JoinSource(JoinRelation.RIGHT);
    ICompute r = new JoinSinkTask();
    computeGraphBuilder.addSource(SOURCE, source1, sourceParallelism);
    computeGraphBuilder.addSource(SOURCE2, source2, sourceParallelism);
    computeConnection = computeGraphBuilder.addCompute(SINK, r, sinkParallelism);
    computeConnection.innerJoin(SOURCE, SOURCE2, CommunicationContext.JoinAlgorithm.SORT).viaLeftEdge(LEFT_EDGE).viaRightEdge(RIGHT_EDGE).withKeyType(keyType).withLeftDataType(dataType).withRightDataType(dataType).withTaskPartitioner(new TaskPartitioner() {

        private List<Integer> dst;

        @Override
        public void prepare(Set sources, Set destinations) {
            this.dst = new ArrayList<>(destinations);
            Collections.sort(this.dst);
        }

        @Override
        public int partition(int source, Object data) {
            return dst.get((Integer) data % dst.size());
        }

        @Override
        public void commit(int source, int partition) {
        }
    }).withComparator(Integer::compareTo);
    return computeGraphBuilder;
}