Examples with WorkerSchedulePlan edu.iu.dsc.tws.api.compute.schedule.elements.WorkerSchedulePlan used on opensource projects

Example 1 with WorkerSchedulePlan

use of edu.iu.dsc.tws.api.compute.schedule.elements.WorkerSchedulePlan 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 2 with WorkerSchedulePlan

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

the class RoundRobinBatchTaskScheduler method schedule.

/**
 * This is the base method which receives the dataflow taskgraph and the worker plan to allocate
 * the task instances to the appropriate workers with their required ram, disk, and cpu values.
 */
@Override
public TaskSchedulePlan schedule(ComputeGraph computeGraph, WorkerPlan workerPlan) {
    Map<Integer, List<TaskInstanceId>> containerInstanceMap;
    Map<Integer, WorkerSchedulePlan> containerPlans = new LinkedHashMap<>();
    for (int i = 0; i < workerPlan.getNumberOfWorkers(); i++) {
        roundRobinAllocation.put(i, new ArrayList<>());
    }
    // To retrieve the batch task instances(it may be single task vertex or a batch of task vertices)
    Set<Vertex> taskVertexSet = new LinkedHashSet<>(computeGraph.getTaskVertexSet());
    TaskVertexParser taskGraphParser = new TaskVertexParser();
    List<Set<Vertex>> taskVertexList = taskGraphParser.parseVertexSet(computeGraph);
    for (Set<Vertex> vertexSet : taskVertexList) {
        if (vertexSet.size() > 1) {
            containerInstanceMap = roundRobinBatchSchedulingAlgorithm(computeGraph, vertexSet);
        } else {
            Vertex vertex = vertexSet.iterator().next();
            containerInstanceMap = roundRobinBatchSchedulingAlgorithm(computeGraph, vertex);
        }
        TaskInstanceMapCalculation instanceMapCalculation = new TaskInstanceMapCalculation(this.instanceRAM, this.instanceCPU, this.instanceDisk);
        Map<Integer, Map<TaskInstanceId, Double>> instancesRamMap = instanceMapCalculation.getInstancesRamMapInContainer(containerInstanceMap, taskVertexSet);
        Map<Integer, Map<TaskInstanceId, Double>> instancesDiskMap = instanceMapCalculation.getInstancesDiskMapInContainer(containerInstanceMap, taskVertexSet);
        Map<Integer, Map<TaskInstanceId, Double>> instancesCPUMap = instanceMapCalculation.getInstancesCPUMapInContainer(containerInstanceMap, taskVertexSet);
        for (int containerId : containerInstanceMap.keySet()) {
            double containerRAMValue = TaskSchedulerContext.containerRamPadding(config);
            double containerDiskValue = TaskSchedulerContext.containerDiskPadding(config);
            double containerCpuValue = TaskSchedulerContext.containerCpuPadding(config);
            List<TaskInstanceId> taskTaskInstanceIds = containerInstanceMap.get(containerId);
            Map<TaskInstanceId, TaskInstancePlan> taskInstancePlanMap = new HashMap<>();
            for (TaskInstanceId id : taskTaskInstanceIds) {
                double instanceRAMValue = instancesRamMap.get(containerId).get(id);
                double instanceDiskValue = instancesDiskMap.get(containerId).get(id);
                double instanceCPUValue = instancesCPUMap.get(containerId).get(id);
                Resource instanceResource = new Resource(instanceRAMValue, instanceDiskValue, instanceCPUValue);
                taskInstancePlanMap.put(id, new TaskInstancePlan(id.getTaskName(), id.getTaskId(), id.getTaskIndex(), instanceResource));
                containerRAMValue += instanceRAMValue;
                containerDiskValue += instanceDiskValue;
                containerCpuValue += instanceDiskValue;
            }
            Worker worker = workerPlan.getWorker(containerId);
            Resource containerResource;
            if (worker != null && worker.getCpu() > 0 && worker.getDisk() > 0 && worker.getRam() > 0) {
                containerResource = new Resource((double) worker.getRam(), (double) worker.getDisk(), (double) worker.getCpu());
            } else {
                containerResource = new Resource(containerRAMValue, containerDiskValue, containerCpuValue);
            }
            WorkerSchedulePlan taskWorkerSchedulePlan;
            if (containerPlans.containsKey(containerId)) {
                taskWorkerSchedulePlan = containerPlans.get(containerId);
                taskWorkerSchedulePlan.getTaskInstances().addAll(taskInstancePlanMap.values());
            } else {
                taskWorkerSchedulePlan = new WorkerSchedulePlan(containerId, new HashSet<>(taskInstancePlanMap.values()), containerResource);
                containerPlans.put(containerId, taskWorkerSchedulePlan);
            }
        }
    }
    return new TaskSchedulePlan(0, new HashSet<>(containerPlans.values()));
}

Example 3 with WorkerSchedulePlan

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

the class DataLocalityStreamingTaskScheduler method schedule.

/**
 * This is the base method for the data locality aware task scheduling for scheduling the
 * streaming task instances. It retrieves the task vertex set of the task graph and send the set
 * to the data locality aware scheduling algorithm to schedule the streaming task instances
 * which are closer to the data nodes.
 */
@Override
public TaskSchedulePlan schedule(ComputeGraph graph, WorkerPlan workerPlan) {
    // Represents task schedule plan Id
    int taskSchedulePlanId = 0;
    Set<WorkerSchedulePlan> workerSchedulePlans = new HashSet<>();
    Set<Vertex> taskVertexSet = graph.getTaskVertexSet();
    Map<Integer, List<TaskInstanceId>> containerInstanceMap = dataLocalityStreamingSchedulingAlgorithm(graph, workerPlan.getNumberOfWorkers(), workerPlan);
    TaskInstanceMapCalculation instanceMapCalculation = new TaskInstanceMapCalculation(this.instanceRAM, this.instanceCPU, this.instanceDisk);
    Map<Integer, Map<TaskInstanceId, Double>> instancesRamMap = instanceMapCalculation.getInstancesRamMapInContainer(containerInstanceMap, taskVertexSet);
    Map<Integer, Map<TaskInstanceId, Double>> instancesDiskMap = instanceMapCalculation.getInstancesDiskMapInContainer(containerInstanceMap, taskVertexSet);
    Map<Integer, Map<TaskInstanceId, Double>> instancesCPUMap = instanceMapCalculation.getInstancesCPUMapInContainer(containerInstanceMap, taskVertexSet);
    for (int containerId : containerInstanceMap.keySet()) {
        double containerRAMValue = TaskSchedulerContext.containerRamPadding(config);
        double containerDiskValue = TaskSchedulerContext.containerDiskPadding(config);
        double containerCpuValue = TaskSchedulerContext.containerCpuPadding(config);
        List<TaskInstanceId> taskTaskInstanceIds = containerInstanceMap.get(containerId);
        Map<TaskInstanceId, TaskInstancePlan> taskInstancePlanMap = new HashMap<>();
        for (TaskInstanceId id : taskTaskInstanceIds) {
            double instanceRAMValue = instancesRamMap.get(containerId).get(id);
            double instanceDiskValue = instancesDiskMap.get(containerId).get(id);
            double instanceCPUValue = instancesCPUMap.get(containerId).get(id);
            Resource instanceResource = new Resource(instanceRAMValue, instanceDiskValue, instanceCPUValue);
            taskInstancePlanMap.put(id, new TaskInstancePlan(id.getTaskName(), id.getTaskId(), id.getTaskIndex(), instanceResource));
            containerRAMValue += instanceRAMValue;
            containerDiskValue += instanceDiskValue;
            containerCpuValue += instanceDiskValue;
        }
        Worker worker = workerPlan.getWorker(containerId);
        Resource containerResource;
        if (worker != null && worker.getCpu() > 0 && worker.getDisk() > 0 && worker.getRam() > 0) {
            containerResource = new Resource((double) worker.getRam(), (double) worker.getDisk(), (double) worker.getCpu());
        } else {
            containerResource = new Resource(containerRAMValue, containerDiskValue, containerCpuValue);
        }
        WorkerSchedulePlan taskWorkerSchedulePlan = new WorkerSchedulePlan(containerId, new HashSet<>(taskInstancePlanMap.values()), containerResource);
        workerSchedulePlans.add(taskWorkerSchedulePlan);
    }
    return new TaskSchedulePlan(taskSchedulePlanId, workerSchedulePlans);
}

Example 4 with WorkerSchedulePlan

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

the class UserDefinedTaskScheduler method schedule.

/**
 * This is the base method which receives the dataflow taskgraph and the worker plan to allocate
 * the task instances to the appropriate workers with their required ram, disk, and cpu values.
 *
 * @return TaskSchedulePlan
 */
@Override
public TaskSchedulePlan schedule(ComputeGraph graph, WorkerPlan workerPlan) {
    int taskSchedulePlanId = 0;
    // Allocate the task instances into the containers/workers
    Set<WorkerSchedulePlan> workerSchedulePlans = new LinkedHashSet<>();
    // To get the vertex set from the taskgraph
    Set<Vertex> taskVertexSet = graph.getTaskVertexSet();
    // Allocate the task instances into the logical containers.
    Map<Integer, List<TaskInstanceId>> userDefinedContainerInstanceMap = userDefinedSchedulingAlgorithm(graph, workerPlan.getNumberOfWorkers());
    TaskInstanceMapCalculation instanceMapCalculation = new TaskInstanceMapCalculation(this.instanceRAM, this.instanceCPU, this.instanceDisk);
    Map<Integer, Map<TaskInstanceId, Double>> instancesRamMap = instanceMapCalculation.getInstancesRamMapInContainer(userDefinedContainerInstanceMap, taskVertexSet);
    Map<Integer, Map<TaskInstanceId, Double>> instancesDiskMap = instanceMapCalculation.getInstancesDiskMapInContainer(userDefinedContainerInstanceMap, taskVertexSet);
    Map<Integer, Map<TaskInstanceId, Double>> instancesCPUMap = instanceMapCalculation.getInstancesCPUMapInContainer(userDefinedContainerInstanceMap, taskVertexSet);
    for (int containerId : userDefinedContainerInstanceMap.keySet()) {
        double containerRAMValue = TaskSchedulerContext.containerRamPadding(config);
        double containerDiskValue = TaskSchedulerContext.containerDiskPadding(config);
        double containerCpuValue = TaskSchedulerContext.containerCpuPadding(config);
        List<TaskInstanceId> taskTaskInstanceIds = userDefinedContainerInstanceMap.get(containerId);
        Map<TaskInstanceId, TaskInstancePlan> taskInstancePlanMap = new HashMap<>();
        for (TaskInstanceId id : taskTaskInstanceIds) {
            double instanceRAMValue = instancesRamMap.get(containerId).get(id);
            double instanceDiskValue = instancesDiskMap.get(containerId).get(id);
            double instanceCPUValue = instancesCPUMap.get(containerId).get(id);
            Resource instanceResource = new Resource(instanceRAMValue, instanceDiskValue, instanceCPUValue);
            taskInstancePlanMap.put(id, new TaskInstancePlan(id.getTaskName(), id.getTaskId(), id.getTaskIndex(), instanceResource));
            containerRAMValue += instanceRAMValue;
            containerDiskValue += instanceDiskValue;
            containerCpuValue += instanceDiskValue;
        }
        Worker worker = workerPlan.getWorker(containerId);
        Resource containerResource;
        // Create the container resource value based on the worker plan
        if (worker != null && worker.getCpu() > 0 && worker.getDisk() > 0 && worker.getRam() > 0) {
            containerResource = new Resource((double) worker.getRam(), (double) worker.getDisk(), (double) worker.getCpu());
            LOG.fine("Worker (if loop):" + containerId + "\tRam:" + worker.getRam() + "\tDisk:" + worker.getDisk() + "\tCpu:" + worker.getCpu());
        } else {
            containerResource = new Resource(containerRAMValue, containerDiskValue, containerCpuValue);
            LOG.fine("Worker (else loop):" + containerId + "\tRam:" + containerRAMValue + "\tDisk:" + containerDiskValue + "\tCpu:" + containerCpuValue);
        }
        // Schedule the task instance plan into the task container plan.
        WorkerSchedulePlan taskWorkerSchedulePlan = new WorkerSchedulePlan(containerId, new HashSet<>(taskInstancePlanMap.values()), containerResource);
        workerSchedulePlans.add(taskWorkerSchedulePlan);
    }
    return new TaskSchedulePlan(taskSchedulePlanId, workerSchedulePlans);
}

Example 5 with WorkerSchedulePlan

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

the class TaskSchedulePlanBuilder method getContainers.

/**
 * Get the containers based on the task schedule plan
 */
private Map<Integer, Container> getContainers(TaskSchedulePlan previoustaskschedulePlan) throws TaskSchedulerException {
    Map<Integer, Container> containerMap = new HashMap<>();
    Resource resource = previoustaskschedulePlan.getMaxContainerResources();
    for (WorkerSchedulePlan currentWorkerSchedulePlan : previoustaskschedulePlan.getContainers()) {
        Container container = new Container(currentWorkerSchedulePlan.getContainerId(), resource, requestedContainerPadding);
        for (TaskInstancePlan instancePlan : currentWorkerSchedulePlan.getTaskInstances()) {
            try {
                addToContainer(container, instancePlan, taskIndexes, taskIds);
            } catch (TaskSchedulerException e) {
                throw new TaskSchedulerException(String.format("Insufficient container resources to add instancePlan %s to container %s", instancePlan, container), e);
            }
        }
        containerMap.put(currentWorkerSchedulePlan.getContainerId(), container);
    }
    return containerMap;
}