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

Example 11 with ExecutionPlan

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

the class SvmSgdAdvancedRunner method executeTestingTaskGraph.

/**
 * This method executes the testing taskgraph with testing data loaded from testing taskgraph
 * and uses the final weight vector obtained from the training task graph
 * Testing is also done in a parallel way. At the testing data loading stage we load the data
 * in parallel with reference to the given parallelism and testing is also in in parallel
 * Then we get test results for all these testing data partitions
 *
 * @return Returns the Accuracy value obtained
 */
public DataObject<Object> executeTestingTaskGraph() {
    DataObject<Object> data = null;
    predictionSourceTask = new PredictionSourceTask(svmJobParameters.isDummy(), this.binaryBatchModel, operationMode);
    predictionReduceTask = new PredictionReduceTask(operationMode);
    testingBuilder.addSource(Constants.SimpleGraphConfig.PREDICTION_SOURCE_TASK, predictionSourceTask, dataStreamerParallelism);
    ComputeConnection predictionReduceConnection = testingBuilder.addCompute(Constants.SimpleGraphConfig.PREDICTION_REDUCE_TASK, predictionReduceTask, reduceParallelism);
    predictionReduceConnection.reduce(Constants.SimpleGraphConfig.PREDICTION_SOURCE_TASK).viaEdge(Constants.SimpleGraphConfig.PREDICTION_EDGE).withReductionFunction(new PredictionAggregator()).withDataType(MessageTypes.OBJECT);
    testingBuilder.setMode(operationMode);
    ComputeGraph predictionGraph = testingBuilder.build();
    predictionGraph.setGraphName("testing-graph");
    ExecutionPlan predictionPlan = taskExecutor.plan(predictionGraph);
    // adding test data set
    taskExecutor.addInput(predictionGraph, predictionPlan, Constants.SimpleGraphConfig.PREDICTION_SOURCE_TASK, Constants.SimpleGraphConfig.TEST_DATA, testingData);
    // adding final weight vector
    taskExecutor.addInput(predictionGraph, predictionPlan, Constants.SimpleGraphConfig.PREDICTION_SOURCE_TASK, Constants.SimpleGraphConfig.FINAL_WEIGHT_VECTOR, trainedWeightVector);
    taskExecutor.execute(predictionGraph, predictionPlan);
    data = retrieveTestingAccuracyObject(predictionGraph, predictionPlan);
    return data;
}

Example 12 with ExecutionPlan

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

the class ExecutionPlanBuilder method build.

@Override
public ExecutionPlan build(Config cfg, ComputeGraph taskGraph, TaskSchedulePlan taskSchedule) {
    // we need to build the task plan
    LogicalPlan logicalPlan = TaskPlanBuilder.build(workerId, workerInfoList, taskSchedule, taskIdGenerator);
    ParallelOperationFactory opFactory = new ParallelOperationFactory(cfg, network, logicalPlan);
    Map<Integer, WorkerSchedulePlan> containersMap = taskSchedule.getContainersMap();
    WorkerSchedulePlan conPlan = containersMap.get(workerId);
    if (conPlan == null) {
        LOG.log(Level.INFO, "Cannot find worker in the task plan: " + workerId);
        return null;
    }
    ExecutionPlan execution = new ExecutionPlan();
    Set<TaskInstancePlan> instancePlan = conPlan.getTaskInstances();
    long tasksVersion = 0L;
    if (CheckpointingContext.isCheckpointingEnabled(cfg)) {
        Set<Integer> globalTasks = Collections.emptySet();
        if (workerId == 0) {
            globalTasks = containersMap.values().stream().flatMap(containerPlan -> containerPlan.getTaskInstances().stream()).filter(ip -> taskGraph.vertex(ip.getTaskName()).getTask() instanceof CheckpointableTask && !(taskGraph.vertex(ip.getTaskName()).getTask() instanceof CheckpointingSGatherSink)).map(TaskInstancePlan::getTaskId).collect(Collectors.toSet());
        }
        try {
            Checkpoint.FamilyInitializeResponse familyInitializeResponse = this.checkpointingClient.initFamily(workerId, containersMap.size(), taskGraph.getGraphName(), globalTasks);
            tasksVersion = familyInitializeResponse.getVersion();
        } catch (BlockingSendException e) {
            throw new RuntimeException("Failed to register tasks with Checkpoint Manager", e);
        }
        LOG.info("Tasks will start with version " + tasksVersion);
    }
    // for each task we are going to create the communications
    for (TaskInstancePlan ip : instancePlan) {
        Vertex v = taskGraph.vertex(ip.getTaskName());
        Map<String, Set<String>> inEdges = new HashMap<>();
        Map<String, String> outEdges = new HashMap<>();
        if (v == null) {
            throw new RuntimeException("Non-existing task scheduled: " + ip.getTaskName());
        }
        INode node = v.getTask();
        if (node instanceof ICompute || node instanceof ISource) {
            // lets get the communication
            Set<Edge> edges = taskGraph.outEdges(v);
            // now lets create the communication object
            for (Edge e : edges) {
                Vertex child = taskGraph.childOfTask(v, e.getName());
                // lets figure out the parents task id
                Set<Integer> srcTasks = taskIdGenerator.getTaskIds(v, ip.getTaskId());
                Set<Integer> tarTasks = taskIdGenerator.getTaskIds(child, getTaskIdOfTask(child.getName(), taskSchedule));
                Map<Integer, Integer> srcGlobalToIndex = taskIdGenerator.getGlobalTaskToIndex(v, ip.getTaskId());
                Map<Integer, Integer> tarGlobaToIndex = taskIdGenerator.getGlobalTaskToIndex(child, getTaskIdOfTask(child.getName(), taskSchedule));
                createCommunication(child, e, v, srcTasks, tarTasks, srcGlobalToIndex, tarGlobaToIndex);
                outEdges.put(e.getName(), child.getName());
            }
        }
        if (node instanceof ICompute) {
            // lets get the parent tasks
            Set<Edge> parentEdges = taskGraph.inEdges(v);
            for (Edge e : parentEdges) {
                Vertex parent = taskGraph.getParentOfTask(v, e.getName());
                // lets figure out the parents task id
                Set<Integer> srcTasks = taskIdGenerator.getTaskIds(parent, getTaskIdOfTask(parent.getName(), taskSchedule));
                Set<Integer> tarTasks = taskIdGenerator.getTaskIds(v, ip.getTaskId());
                Map<Integer, Integer> srcGlobalToIndex = taskIdGenerator.getGlobalTaskToIndex(parent, getTaskIdOfTask(parent.getName(), taskSchedule));
                Map<Integer, Integer> tarGlobalToIndex = taskIdGenerator.getGlobalTaskToIndex(v, ip.getTaskId());
                createCommunication(v, e, parent, srcTasks, tarTasks, srcGlobalToIndex, tarGlobalToIndex);
                // if we are a grouped edge, we have to use the group name
                String inEdge;
                if (e.getTargetEdge() == null) {
                    inEdge = e.getName();
                } else {
                    inEdge = e.getTargetEdge();
                }
                Set<String> parents = inEdges.get(inEdge);
                if (parents == null) {
                    parents = new HashSet<>();
                }
                parents.add(inEdge);
                inEdges.put(inEdge, parents);
            }
        }
        // lets create the instance
        INodeInstance iNodeInstance = createInstances(cfg, taskGraph.getGraphName(), ip, v, taskGraph.getOperationMode(), inEdges, outEdges, taskSchedule, tasksVersion);
        // add to execution
        execution.addNodes(v.getName(), taskIdGenerator.generateGlobalTaskId(ip.getTaskId(), ip.getTaskIndex()), iNodeInstance);
    }
    // now lets create the queues and start the execution
    for (Table.Cell<String, String, Communication> cell : parOpTable.cellSet()) {
        Communication c = cell.getValue();
        // lets create the communication
        OperationMode operationMode = taskGraph.getOperationMode();
        IParallelOperation op;
        assert c != null;
        c.build();
        if (c.getEdge().size() == 1) {
            op = opFactory.build(c.getEdge(0), c.getSourceTasks(), c.getTargetTasks(), operationMode, c.srcGlobalToIndex, c.tarGlobalToIndex);
        } else if (c.getEdge().size() > 1) {
            // just join op for now. Could change in the future
            // here the sources should be separated out for left and right edge
            Set<Integer> sourceTasks = c.getSourceTasks();
            Set<Integer> leftSources = new HashSet<>();
            Set<Integer> rightSources = new HashSet<>();
            if (!sourceTasks.isEmpty()) {
                // just to safely do .get() calls without isPresent()
                int minBin = (sourceTasks.stream().min(Integer::compareTo).get() / TaskIdGenerator.TASK_OFFSET) * TaskIdGenerator.TASK_OFFSET;
                for (Integer source : sourceTasks) {
                    if ((source / TaskIdGenerator.TASK_OFFSET) * TaskIdGenerator.TASK_OFFSET == minBin) {
                        leftSources.add(source);
                    } else {
                        rightSources.add(source);
                    }
                }
            }
            // now determine, which task is connected to which edge
            Edge leftEdge = c.getEdge(0);
            Edge rightEdge = c.getEdge(1);
            op = opFactory.build(leftEdge, rightEdge, leftSources, rightSources, c.getTargetTasks(), operationMode, c.srcGlobalToIndex, c.tarGlobalToIndex);
        } else {
            throw new RuntimeException("Cannot have communication with 0 edges");
        }
        // now lets check the sources and targets that are in this executor
        Set<Integer> sourcesOfThisWorker = intersectionOfTasks(conPlan, c.getSourceTasks());
        Set<Integer> targetsOfThisWorker = intersectionOfTasks(conPlan, c.getTargetTasks());
        // we use the target edge as the group name
        String targetEdge;
        if (c.getEdge().size() > 1) {
            targetEdge = c.getEdge(0).getTargetEdge();
        } else {
            targetEdge = c.getEdge(0).getName();
        }
        // so along with the operation mode, the windowing mode must be tested
        if (operationMode == OperationMode.STREAMING) {
            for (Integer i : sourcesOfThisWorker) {
                boolean found = false;
                // we can have multiple source tasks for an operation
                for (int sIndex = 0; sIndex < c.getSourceTask().size(); sIndex++) {
                    String sourceTask = c.getSourceTask().get(sIndex);
                    if (streamingTaskInstances.contains(sourceTask, i)) {
                        TaskStreamingInstance taskStreamingInstance = streamingTaskInstances.get(sourceTask, i);
                        taskStreamingInstance.registerOutParallelOperation(c.getEdge(sIndex).getName(), op);
                        op.registerSync(i, taskStreamingInstance);
                        found = true;
                    } else if (streamingSourceInstances.contains(sourceTask, i)) {
                        SourceStreamingInstance sourceStreamingInstance = streamingSourceInstances.get(sourceTask, i);
                        sourceStreamingInstance.registerOutParallelOperation(c.getEdge(sIndex).getName(), op);
                        found = true;
                    }
                    if (!found) {
                        throw new RuntimeException("Not found: " + c.getSourceTask());
                    }
                }
            }
            // we only have one target task always
            for (Integer i : targetsOfThisWorker) {
                if (streamingTaskInstances.contains(c.getTargetTask(), i)) {
                    TaskStreamingInstance taskStreamingInstance = streamingTaskInstances.get(c.getTargetTask(), i);
                    op.register(i, taskStreamingInstance.getInQueue());
                    taskStreamingInstance.registerInParallelOperation(targetEdge, op);
                    op.registerSync(i, taskStreamingInstance);
                } else {
                    throw new RuntimeException("Not found: " + c.getTargetTask());
                }
            }
            execution.addOps(op);
        }
        if (operationMode == OperationMode.BATCH) {
            for (Integer i : sourcesOfThisWorker) {
                boolean found = false;
                // we can have multiple source tasks for an operation
                for (int sIndex = 0; sIndex < c.getSourceTask().size(); sIndex++) {
                    String sourceTask = c.getSourceTask().get(sIndex);
                    if (batchTaskInstances.contains(sourceTask, i)) {
                        TaskBatchInstance taskBatchInstance = batchTaskInstances.get(sourceTask, i);
                        taskBatchInstance.registerOutParallelOperation(c.getEdge(sIndex).getName(), op);
                        found = true;
                    } else if (batchSourceInstances.contains(sourceTask, i)) {
                        SourceBatchInstance sourceBatchInstance = batchSourceInstances.get(sourceTask, i);
                        sourceBatchInstance.registerOutParallelOperation(c.getEdge(sIndex).getName(), op);
                        found = true;
                    }
                }
                if (!found) {
                    throw new RuntimeException("Not found: " + c.getSourceTask());
                }
            }
            for (Integer i : targetsOfThisWorker) {
                if (batchTaskInstances.contains(c.getTargetTask(), i)) {
                    TaskBatchInstance taskBatchInstance = batchTaskInstances.get(c.getTargetTask(), i);
                    op.register(i, taskBatchInstance.getInQueue());
                    taskBatchInstance.registerInParallelOperation(targetEdge, op);
                    op.registerSync(i, taskBatchInstance);
                } else {
                    throw new RuntimeException("Not found: " + c.getTargetTask());
                }
            }
            execution.addOps(op);
        }
    }
    return execution;
}

Example 13 with ExecutionPlan

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

the class BasicComputation method execute.

@Override
public void execute() {
    LOG.log(Level.INFO, "Task worker starting: " + workerId);
    WorkerParameter workerParameter = WorkerParameter.build(config);
    parallelism = workerParameter.getParallelismValue();
    graphsize = workerParameter.getDsize();
    sourceVertexGlobal = workerParameter.getSourcevertex();
    dataDirectory = workerParameter.getDatapointDirectory();
    iterations = workerParameter.getIterations();
    /* First Graph to partition and read the partitioned data points **/
    ComputeGraph graphpartitionTaskGraph = graphpartition();
    // Get the execution plan for the first task graph
    ExecutionPlan executionPlan = taskExecutor.plan(graphpartitionTaskGraph);
    // Actual execution for the first taskgraph
    taskExecutor.execute(graphpartitionTaskGraph, executionPlan);
    // Retrieve the output of the first task graph
    // +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
    // the second task graph for assign initial pagerank values for vertex.
    ComputeGraph graphInitializationTaskGraph = graphInitialization();
    // Get the execution plan for the first task graph
    ExecutionPlan executionPlan1 = taskExecutor.plan(graphInitializationTaskGraph);
    // Actual execution for the first taskgraph
    taskExecutor.execute(graphInitializationTaskGraph, executionPlan1);
    // Retrieve the output of the first task graph
    // +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
    // third task graph for computations
    ComputeGraph computationTaskgraph = computation();
    IExecutor ex = taskExecutor.createExecution(computationTaskgraph);
    int unlimitedItr = 0;
    long startime = System.currentTimeMillis();
    if (iterations != 0) {
        for (int i = 0; i < iterations; i++) {
            ex.execute(i == iterations - 1);
        }
    } else {
        while (globaliterationStatus) {
            ex.execute(false);
            unlimitedItr++;
        }
        ex.closeExecution();
    }
    taskExecutor.close();
    long endTime = System.currentTimeMillis();
    if (workerId == 0) {
        System.out.println("total number of iterations : " + unlimitedItr);
        System.out.println("computation time: " + (endTime - startime));
    }
}

Example 14 with ExecutionPlan

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

the class PageRankWorker method execute.

// private static double danglingNodeValues;
@Override
public void execute() {
    LOG.log(Level.INFO, "Task worker starting: " + workerId);
    PageRankWorkerParameters pageRankWorkerParameters = PageRankWorkerParameters.build(config);
    int parallelismValue = pageRankWorkerParameters.getParallelismValue();
    int dsize = pageRankWorkerParameters.getDsize();
    String dataDirectory = pageRankWorkerParameters.getDatapointDirectory();
    int iterations = pageRankWorkerParameters.getIterations();
    graphsize = dsize;
    /* First Graph to partition and read the partitioned data points **/
    ComputeGraph datapointsTaskGraph = buildDataPointsTG(dataDirectory, dsize, parallelismValue, config);
    // Get the execution plan for the first task graph
    ExecutionPlan executionPlan = taskExecutor.plan(datapointsTaskGraph);
    // Actual execution for the first taskgraph
    taskExecutor.execute(datapointsTaskGraph, executionPlan);
    /* the out of the first graph would like below
    * task Id: 0
    {1=[3, 4], 2=[3, 4, 5]}*/
    // +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
    // the second task graph for assign initial pagerank values for vertex.
    ComputeGraph graphInitialValueTaskGraph = buildGraphInitialValueTG(dataDirectory, dsize, parallelismValue, config);
    // Get the execution plan for the first task graph
    ExecutionPlan executionPlan1 = taskExecutor.plan(graphInitialValueTaskGraph);
    // Actual execution for the first taskgraph
    taskExecutor.execute(graphInitialValueTaskGraph, executionPlan1);
    /* the output of second graph should like below
      initiate the pagerank value
    * {1=0.25, 2=0.25}
     */
    // +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
    // third task graph for computations
    ComputeGraph pageranktaskgraph = buildComputationTG(parallelismValue, config);
    IExecutor ex = taskExecutor.createExecution(pageranktaskgraph);
    // Perform the iterations from 0 to 'n' number of iterations
    long startime = System.currentTimeMillis();
    for (int i = 0; i < iterations; i++) {
        ex.execute(i == iterations - 1);
    }
    taskExecutor.close();
    long endTime = System.currentTimeMillis();
    if (workerId == 0) {
        DataObject<Object> graphInitialPagerankValue = taskExecutor.getOutput("InitialValue");
        DataPartition<?> finaloutput = graphInitialPagerankValue.getPartition(workerId);
        HashMap<String, Double> finalone = (HashMap<String, Double>) finaloutput.getConsumer().next();
        System.out.println(finalone);
        LOG.info("Final output After " + iterations + "iterations ");
        Iterator it = finalone.entrySet().iterator();
        Double recivedFinalDanglingValue = finalone.get("danglingvalues");
        double cummulativepagerankvalue = 0.0;
        int num = 0;
        System.out.println(graphsize);
        while (it.hasNext()) {
            Map.Entry pair = (Map.Entry) it.next();
            if (!pair.getKey().equals("danglingvalues")) {
                double finalPagerankValue = (double) pair.getValue() + ((0.85 * recivedFinalDanglingValue) / graphsize);
                System.out.print("Vertex Id: " + pair.getKey());
                System.out.printf(" and it's pagerank value: %.15f \n", finalPagerankValue);
                cummulativepagerankvalue += finalPagerankValue;
                num += 1;
            }
            // avoids a ConcurrentModificationException
            it.remove();
        }
        System.out.println(recivedFinalDanglingValue);
        System.out.println(num);
        System.out.println(cummulativepagerankvalue);
        System.out.println(cummulativepagerankvalue + ((graphsize - num) * ((((double) 1 / graphsize) * 0.15) + (0.85 * (recivedFinalDanglingValue / graphsize)))));
        System.out.println("computation time: " + (endTime - startime));
    }
}

Example 15 with ExecutionPlan

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

the class Twister2StormWorker method execute.

@Override
public void execute() {
    ComputeGraph graph = this.buildTopology().getT2ComputeGraph();
    ExecutionPlan executionPlan = taskExecutor.plan(graph);
    taskExecutor.execute(graph, executionPlan);
}