use of edu.iu.dsc.tws.api.compute.schedule.elements.TaskInstancePlan in project twister2 by DSC-SPIDAL.
the class BatchTaskScheduler 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.
*
* @param computeGraph
* @param workerPlan worker plan
* @return
*/
@Override
public TaskSchedulePlan schedule(ComputeGraph computeGraph, WorkerPlan workerPlan) {
// Allocate the task instances into the containers/workers
Set<WorkerSchedulePlan> workerSchedulePlans = new LinkedHashSet<>();
// To get the vertex set from the Collectible Name Settaskgraph
Set<Vertex> taskVertexSet = new LinkedHashSet<>(computeGraph.getTaskVertexSet());
// Allocate the task instances into the logical containers.
Map<Integer, List<TaskInstanceId>> batchContainerInstanceMap = batchSchedulingAlgorithm(computeGraph, workerPlan.getNumberOfWorkers());
TaskInstanceMapCalculation instanceMapCalculation = new TaskInstanceMapCalculation(this.instanceRAM, this.instanceCPU, this.instanceDisk);
Map<Integer, Map<TaskInstanceId, Double>> instancesRamMap = instanceMapCalculation.getInstancesRamMapInContainer(batchContainerInstanceMap, taskVertexSet);
Map<Integer, Map<TaskInstanceId, Double>> instancesDiskMap = instanceMapCalculation.getInstancesDiskMapInContainer(batchContainerInstanceMap, taskVertexSet);
Map<Integer, Map<TaskInstanceId, Double>> instancesCPUMap = instanceMapCalculation.getInstancesCPUMapInContainer(batchContainerInstanceMap, taskVertexSet);
for (int containerId : batchContainerInstanceMap.keySet()) {
double containerRAMValue = TaskSchedulerContext.containerRamPadding(config);
double containerDiskValue = TaskSchedulerContext.containerDiskPadding(config);
double containerCpuValue = TaskSchedulerContext.containerCpuPadding(config);
List<TaskInstanceId> taskTaskInstanceIds = batchContainerInstanceMap.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());
} else {
containerResource = new Resource(containerRAMValue, containerDiskValue, containerCpuValue);
}
// Schedule the task instance plan into the task container plan.
WorkerSchedulePlan taskWorkerSchedulePlan = new WorkerSchedulePlan(containerId, new LinkedHashSet<>(taskInstancePlanMap.values()), containerResource);
workerSchedulePlans.add(taskWorkerSchedulePlan);
if (dependentGraphs && index == 0) {
workerIdList.add(containerId);
}
}
index++;
TaskSchedulePlan taskSchedulePlan = new TaskSchedulePlan(0, workerSchedulePlans);
if (workerId == 0) {
Map<Integer, WorkerSchedulePlan> containersMap = taskSchedulePlan.getContainersMap();
for (Map.Entry<Integer, WorkerSchedulePlan> entry : containersMap.entrySet()) {
Integer integer = entry.getKey();
WorkerSchedulePlan workerSchedulePlan = entry.getValue();
Set<TaskInstancePlan> containerPlanTaskInstances = workerSchedulePlan.getTaskInstances();
LOG.fine("Graph Name:" + computeGraph.getGraphName() + "\tcontainer id:" + integer);
for (TaskInstancePlan ip : containerPlanTaskInstances) {
LOG.fine("Task Id:" + ip.getTaskId() + "\tIndex" + ip.getTaskIndex() + "\tName:" + ip.getTaskName());
}
}
}
return taskSchedulePlan;
}
use of edu.iu.dsc.tws.api.compute.schedule.elements.TaskInstancePlan in project twister2 by DSC-SPIDAL.
the class TaskPlanBuilder method build.
/**
* Create a task plan based on the resource plan from resources and scheduled plan
*
* @param schedulePlan schedule plan
* @param idGenerator global task id generator
* @return the task plan
*/
public static LogicalPlan build(int workerID, List<JobMasterAPI.WorkerInfo> workerInfoList, TaskSchedulePlan schedulePlan, TaskIdGenerator idGenerator) {
Set<WorkerSchedulePlan> cPlanList = schedulePlan.getContainers();
Map<Integer, Set<Integer>> containersToTasks = new HashMap<>();
Map<Integer, Set<Integer>> groupsToTasks = new HashMap<>();
// we need to sort to keep the order
workerInfoList.sort(Comparator.comparingInt(JobMasterAPI.WorkerInfo::getWorkerID));
for (WorkerSchedulePlan c : cPlanList) {
Set<TaskInstancePlan> tSet = c.getTaskInstances();
Set<Integer> instances = new HashSet<>();
for (TaskInstancePlan tPlan : tSet) {
instances.add(idGenerator.generateGlobalTaskId(tPlan.getTaskId(), tPlan.getTaskIndex()));
}
containersToTasks.put(c.getContainerId(), instances);
}
Map<String, List<JobMasterAPI.WorkerInfo>> containersPerNode = new TreeMap<>();
for (JobMasterAPI.WorkerInfo workerInfo : workerInfoList) {
String name = Integer.toString(workerInfo.getWorkerID());
List<JobMasterAPI.WorkerInfo> containerList;
if (!containersPerNode.containsKey(name)) {
containerList = new ArrayList<>();
containersPerNode.put(name, containerList);
} else {
containerList = containersPerNode.get(name);
}
containerList.add(workerInfo);
}
Map<String, Set<Integer>> nodeToTasks = new HashMap<>();
int i = 0;
// we take each container as an executor
for (Map.Entry<String, List<JobMasterAPI.WorkerInfo>> entry : containersPerNode.entrySet()) {
Set<Integer> executorsOfGroup = new HashSet<>();
for (JobMasterAPI.WorkerInfo workerInfo : entry.getValue()) {
executorsOfGroup.add(workerInfo.getWorkerID());
Set<Integer> tasksInNode = nodeToTasks.computeIfAbsent(workerInfo.getNodeInfo().getNodeIP(), k -> new HashSet<>());
tasksInNode.addAll(containersToTasks.getOrDefault(workerInfo.getWorkerID(), Collections.emptySet()));
}
groupsToTasks.put(i, executorsOfGroup);
i++;
}
return new LogicalPlan(containersToTasks, groupsToTasks, nodeToTasks, workerID);
}
use of edu.iu.dsc.tws.api.compute.schedule.elements.TaskInstancePlan in project twister2 by DSC-SPIDAL.
the class RoundRobinTaskSchedulerTest method testUniqueSchedules4.
@Test
public void testUniqueSchedules4() {
int parallel = 16;
int workers = 2;
ComputeGraph graph = createGraphWithComputeTaskAndConstraints(parallel);
RoundRobinTaskScheduler scheduler = new RoundRobinTaskScheduler();
scheduler.initialize(Config.newBuilder().build());
WorkerPlan workerPlan = createWorkPlan(workers);
TaskSchedulePlan plan1 = scheduler.schedule(graph, workerPlan);
Map<Integer, WorkerSchedulePlan> containersMap = plan1.getContainersMap();
for (Map.Entry<Integer, WorkerSchedulePlan> entry : containersMap.entrySet()) {
WorkerSchedulePlan workerSchedulePlan = entry.getValue();
Set<TaskInstancePlan> containerPlanTaskInstances = workerSchedulePlan.getTaskInstances();
Assert.assertEquals(containerPlanTaskInstances.size(), Integer.parseInt(graph.getGraphConstraints().get(Context.TWISTER2_MAX_TASK_INSTANCES_PER_WORKER)));
}
}
use of edu.iu.dsc.tws.api.compute.schedule.elements.TaskInstancePlan in project twister2 by DSC-SPIDAL.
the class DataLocalityBatchTaskScheduler method schedule.
/**
* This is the base method for the data locality aware task scheduling for scheduling the batch
* task instances. It retrieves the task vertex set of the task graph and send the set to the
* data locality aware scheduling algorithm to allocate the batch task instances which are closer
* to the data nodes.
*/
@Override
public TaskSchedulePlan schedule(ComputeGraph graph, WorkerPlan workerPlan) {
LinkedHashMap<Integer, WorkerSchedulePlan> containerPlans = new LinkedHashMap<>();
for (int i = 0; i < workerPlan.getNumberOfWorkers(); i++) {
dataLocalityAwareAllocation.put(i, new ArrayList<>());
}
LinkedHashSet<Vertex> taskVertexSet = new LinkedHashSet<>(graph.getTaskVertexSet());
TaskVertexParser taskVertexParser = new TaskVertexParser();
List<Set<Vertex>> taskVertexList = taskVertexParser.parseVertexSet(graph);
for (Set<Vertex> vertexSet : taskVertexList) {
Map<Integer, List<TaskInstanceId>> containerInstanceMap;
if (vertexSet.size() > 1) {
containerInstanceMap = dataLocalityBatchSchedulingAlgorithm(graph, vertexSet, workerPlan);
} else {
Vertex vertex = vertexSet.iterator().next();
containerInstanceMap = dataLocalityBatchSchedulingAlgorithm(graph, vertex, 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;
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);
}
}
}
TaskSchedulePlan taskSchedulePlan = new TaskSchedulePlan(0, new HashSet<>(containerPlans.values()));
Map<Integer, WorkerSchedulePlan> containersMap = taskSchedulePlan.getContainersMap();
for (Map.Entry<Integer, WorkerSchedulePlan> entry : containersMap.entrySet()) {
Integer integer = entry.getKey();
WorkerSchedulePlan workerSchedulePlan = entry.getValue();
Set<TaskInstancePlan> containerPlanTaskInstances = workerSchedulePlan.getTaskInstances();
LOG.fine("Task Details for Container Id:" + integer + "\tsize:" + containerPlanTaskInstances.size());
for (TaskInstancePlan ip : containerPlanTaskInstances) {
LOG.fine("TaskId:" + ip.getTaskId() + "\tTask Index" + ip.getTaskIndex() + "\tTask Name:" + ip.getTaskName());
}
}
return taskSchedulePlan;
}
use of edu.iu.dsc.tws.api.compute.schedule.elements.TaskInstancePlan in project twister2 by DSC-SPIDAL.
the class TaskScheduler method generateTaskSchedulePlan.
private TaskSchedulePlan generateTaskSchedulePlan(String className) {
Class<?> taskSchedulerClass;
Method method;
TaskSchedulePlan taskSchedulePlan;
try {
taskSchedulerClass = getClass().getClassLoader().loadClass(className);
Object newInstance = taskSchedulerClass.newInstance();
method = taskSchedulerClass.getMethod("initialize", new Class<?>[] { Config.class });
method.invoke(newInstance, config);
method = taskSchedulerClass.getMethod("schedule", new Class<?>[] { ComputeGraph.class, WorkerPlan.class });
taskSchedulePlan = (TaskSchedulePlan) method.invoke(newInstance, computeGraph, workerPlan);
} catch (InvocationTargetException | IllegalAccessException | NoSuchMethodException | InstantiationException | ClassNotFoundException | TaskSchedulerException e) {
throw new Twister2RuntimeException(e);
}
if (taskSchedulePlan != null) {
Map<Integer, WorkerSchedulePlan> containersMap = taskSchedulePlan.getContainersMap();
for (Map.Entry<Integer, WorkerSchedulePlan> entry : containersMap.entrySet()) {
Integer integer = entry.getKey();
WorkerSchedulePlan workerSchedulePlan = entry.getValue();
Set<TaskInstancePlan> containerPlanTaskInstances = workerSchedulePlan.getTaskInstances();
LOG.fine("Task Details for Container Id:" + integer);
for (TaskInstancePlan ip : containerPlanTaskInstances) {
LOG.fine("Task Id:" + ip.getTaskId() + "\tTask Index" + ip.getTaskIndex() + "\tTask Name:" + ip.getTaskName());
}
}
}
return taskSchedulePlan;
}
Aggregations