Flink – SlotSharingGroup

 

SlotSharingGroup

表示不同的task可以共享slot,但是这是soft的约束,即也可以不在一个slot

默认情况下,整个StreamGraph都会用一个默认的“default” SlotSharingGroup,即所有的JobVertex的task都可以共用一个slot

/**
 * A slot sharing units defines which different task (from different job vertices) can be
 * deployed together within a slot. This is a soft permission, in contrast to the hard constraint
 * defined by a co-location hint.
 */
public class SlotSharingGroup implements java.io.Serializable {
    private final Set<JobVertexID> ids = new TreeSet<JobVertexID>();
    
    /** Mapping of tasks to subslots. This field is only needed inside the JobManager, and is not RPCed. */
    private transient SlotSharingGroupAssignment taskAssignment;

可以看到,关键的部分

1. TreeSet<JobVertexID>(),用于保存属于这个group的Jobvertexid

2. SlotSharingGroupAssignment

 

CoLocationGroup

只是hard约束,在group中的JobVertices,需要对应index的subtask跑在一个slot中

/**
 * A Co-location group is a group of JobVertices, where the <i>i-th</i> subtask of one vertex
 * has to be executed on the same TaskManager as the <i>i-th</i> subtask of all
 * other JobVertices in the same group.
 * 
 * <p>The co-location group is used for example to make sure that the i-th subtasks for iteration
 * head and iteration tail are scheduled to the same TaskManager.</p>
 */
public class CoLocationGroup implements java.io.Serializable {

    /** The ID that describes the slot co-location-constraint as a group */ 
    private final AbstractID id = new AbstractID();
    
    /** The vertices participating in the co-location group */
    private final List<JobVertex> vertices = new ArrayList<JobVertex>();
    
    /** The constraints, which hold the shared slots for the co-located operators */
    private transient ArrayList<CoLocationConstraint> constraints;

 

CoLocationConstraint,可以看作一种特殊的SharedSlot

/**
 * A CoLocationConstraint manages the location of a set of tasks
 * (Execution Vertices). In co-location groups, the different subtasks of
 * different JobVertices need to be executed on the same {@link Instance}.
 * This is realized by creating a special shared slot that holds these tasks.
 * 
 * <p>This class tracks the location and the shared slot for this set of tasks.
 */
public class CoLocationConstraint {

    private final CoLocationGroup group;

    private volatile SharedSlot sharedSlot;

    private volatile ResourceID lockedLocation;

 

几种Slot,

AllocatedSlot ,代表从taskmanager分配出的slot

/**
 * The {@code AllocatedSlot} represents a slot that the JobManager allocated from a TaskManager.
 * It represents a slice of allocated resources from the TaskManager.
 * 
 * <p>To allocate an {@code AllocatedSlot}, the requests a slot from the ResourceManager. The
 * ResourceManager picks (or starts) a TaskManager that will then allocate the slot to the
 * JobManager and notify the JobManager.
 * 
 * <p>Note: Prior to the resource management changes introduced in (Flink Improvement Proposal 6),
 * an AllocatedSlot was allocated to the JobManager as soon as the TaskManager registered at the
 * JobManager. All slots had a default unknown resource profile. 
 */
public class AllocatedSlot {

    /** The ID under which the slot is allocated. Uniquely identifies the slot. */
    private final AllocationID slotAllocationId;

    /** The ID of the job this slot is allocated for */
    private final JobID jobID;

    /** The location information of the TaskManager to which this slot belongs */
    private final TaskManagerLocation taskManagerLocation;

    /** The resource profile of the slot provides */
    private final ResourceProfile resourceProfile;

    /** RPC gateway to call the TaskManager that holds this slot */
    private final TaskManagerGateway taskManagerGateway;

    /** The number of the slot on the TaskManager to which slot belongs. Purely informational. */
    private final int slotNumber;

 

Slot,可以看作对AllocatedSlot的封装

/**
 * Base class for slots that the Scheduler / ExecutionGraph take from the SlotPool and use to place
 * tasks to execute into. A slot corresponds to an AllocatedSlot (a slice of a TaskManager's resources),
 * plus additional fields to track what is currently executed in that slot, or if the slot is still
 * used or disposed (ExecutionGraph gave it back to the pool).
 *
 * <p>In the simplest case, a slot holds a single task ({@link SimpleSlot}). In the more complex
 * case, a slot is shared ({@link SharedSlot}) and contains a set of tasks. Shared slots may contain
 * other shared slots which in turn can hold simple slots. That way, a shared slot may define a tree
 * of slots that belong to it.
 */
public abstract class Slot {

    /** The allocated slot that this slot represents. */
    private final AllocatedSlot allocatedSlot;

    /** The owner of this slot - the slot was taken from that owner and must be disposed to it */
    private final SlotOwner owner;

    /** The parent of this slot in the hierarchy, or null, if this is the parent */
    @Nullable
    private final SharedSlot parent;

    /** The id of the group that this slot is allocated to. May be null. */
    @Nullable
    private final AbstractID groupID;

    /** The number of the slot on which the task is deployed */
    private final int slotNumber;

 

SimpleSlot,放单个task的slot

/**
 * A SimpleSlot represents a single slot on a TaskManager instance, or a slot within a shared slot.
 *
 * <p>If this slot is part of a {@link SharedSlot}, then the parent attribute will point to that shared slot.
 * If not, then the parent attribute is null.
 */
public class SimpleSlot extends Slot {

    /** Task being executed in the slot. Volatile to force a memory barrier and allow for correct double-checking */
    private volatile Execution executedTask; //非share,只有一个task

    /** The locality attached to the slot, defining whether the slot was allocated at the desired location. */
    private volatile Locality locality = Locality.UNCONSTRAINED;

 

SharedSlot ,

/**
 * This class represents a shared slot. A shared slot can have multiple
 * {@link SimpleSlot} instances within itself. This allows to
 * schedule multiple tasks simultaneously to the same resource. Sharing a resource with multiple
 * tasks is crucial for simple pipelined / streamed execution, where both the sender and the receiver
 * are typically active at the same time.
 *
 * <p><b>IMPORTANT:</b> This class contains no synchronization. Thus, the caller has to guarantee proper
 * synchronization. In the current implementation, all concurrently modifying operations are
 * passed through a {@link SlotSharingGroupAssignment} object which is responsible for
 * synchronization.
 */
public class SharedSlot extends Slot {

    /** The assignment group os shared slots that manages the availability and release of the slots */
    private final SlotSharingGroupAssignment assignmentGroup;

    /** The set os sub-slots allocated from this shared slot */
    private final Set<Slot> subSlots;

可以看到sharedSlot继承自Slot,而Slot中只有一个

AllocatedSlot allocatedSlot

所以,无论在subSlots有多少slot,但他们都是共用这个allocatedSlot的

从相应的sharedSlot上去分配simpleSlot

    SimpleSlot allocateSubSlot(AbstractID groupId) {
        if (isAlive()) {
            SimpleSlot slot = new SimpleSlot(
                    getJobID(), getOwner(), getTaskManagerLocation(), subSlots.size(), 
                    getTaskManagerGateway(), this, groupId);
            subSlots.add(slot);
            return slot;
        }
        else {
            return null;
        }
    }

 

    public SimpleSlot(
            JobID jobID, SlotOwner owner, TaskManagerLocation location, int slotNumber,
            TaskManagerGateway taskManagerGateway,
            @Nullable SharedSlot parent, @Nullable AbstractID groupID) {

        super(parent != null ? //如果有parent,即属于某个sharedSlot
                parent.getAllocatedSlot() : //使用parent sharedSlot
                new AllocatedSlot(NO_ALLOCATION_ID, jobID, location, slotNumber, //创建新的AllocatedSlot
                        ResourceProfile.UNKNOWN, taskManagerGateway),
                owner, slotNumber, parent, groupID);
    }

 

 

SlotSharingGroupAssignment,用于管理一组SharedSlot

注释中的示意图,比较清晰

/**
 * The SlotSharingGroupAssignment manages a set of shared slots, which are shared between
 * tasks of a {@link org.apache.flink.runtime.jobmanager.scheduler.SlotSharingGroup}.
 * 
 * <p>The assignments shares tasks by allowing a shared slot to hold one vertex per
 * JobVertexID. For example, consider a program consisting of job vertices "source", "map",
 * "reduce", and "sink". If the slot sharing group spans all four job vertices, then
 * each shared slot can hold one parallel subtask of the source, the map, the reduce, and the
 * sink vertex. Each shared slot holds the actual subtasks in child slots, which are (at the leaf level),
 * the {@link SimpleSlot}s.</p>
 * 
 * <p>An exception are the co-location-constraints, that define that the i-th subtask of one
 * vertex needs to be scheduled strictly together with the i-th subtasks of of the vertices
 * that share the co-location-constraint. To manage that, a co-location-constraint gets its
 * own shared slot inside the shared slots of a sharing group.</p>
 * 
 * <p>Consider a job set up like this:</p>
 * 
 * <pre>{@code
 * +-------------- Slot Sharing Group --------------+
 * |                                                |
 * |            +-- Co Location Group --+           |
 * |            |                       |           |
 * |  (source) ---> (head) ---> (tail) ---> (sink)  |
 * |            |                       |           |
 * |            +-----------------------+           |
 * +------------------------------------------------+
 * }</pre>
 * 
 * <p>The slot hierarchy in the slot sharing group will look like the following</p> 
 * 
 * <pre>
 *     Shared(0)(root)
 *        |
 *        +-- Simple(2)(sink)
 *        |
 *        +-- Shared(1)(co-location-group)
 *        |      |
 *        |      +-- Simple(0)(tail)
 *        |      +-- Simple(1)(head)
 *        |
 *        +-- Simple(0)(source)
 * </pre>
 */
public class SlotSharingGroupAssignment {
    
    /** All slots currently allocated to this sharing group */
    private final Set<SharedSlot> allSlots = new LinkedHashSet<SharedSlot>();

    /** The slots available per vertex type (JobVertexId), keyed by TaskManager, to make them locatable */
    private final Map<AbstractID, Map<ResourceID, List<SharedSlot>>> availableSlotsPerJid = new LinkedHashMap<>();

核心的结构,

allSlots,用于保存所有的SharedSlot,这些SharedSlot都是可以共享的,被分配给不同的JobVertex下的task

availableSlotsPerJid,用于记录对应关系,AbstractID表示JobVertexID,ResourceID表示TaskManager

 

最核心的函数,

getSlotForTask,为task分配slot

   /**
     * Gets a slot suitable for the given task vertex. This method will prefer slots that are local
     * (with respect to {@link ExecutionVertex#getPreferredLocationsBasedOnInputs()}), but will return non local
     * slots if no local slot is available. The method returns null, when this sharing group has
     * no slot is available for the given JobVertexID. 
     *
     * @param vertex The vertex to allocate a slot for.
     *
     * @return A slot to execute the given ExecutionVertex in, or null, if none is available.
     */
    public SimpleSlot getSlotForTask(ExecutionVertex vertex) {
        return getSlotForTask(vertex.getJobvertexId(), vertex.getPreferredLocationsBasedOnInputs()); //默认以input所分配的slot的location信息,作为Preferred
    }

    /**
     * 
     */
    SimpleSlot getSlotForTask(JobVertexID vertexID, Iterable<TaskManagerLocation> locationPreferences) {
        synchronized (lock) {
            Tuple2<SharedSlot, Locality> p = getSlotForTaskInternal(vertexID, locationPreferences, false); //获取SharedSlot,第三个参数意思是,不是一定要local

            if (p != null) {
                SharedSlot ss = p.f0;
                SimpleSlot slot = ss.allocateSubSlot(vertexID); //从SharedSlot中分配SimpleSlot
                slot.setLocality(p.f1);
                return slot;
            }
            else {
                return null;
            }
        }
    }

 

getSlotForTaskInternal

  private Tuple2<SharedSlot, Locality> getSlotForTaskInternal(
            AbstractID groupId, Iterable<TaskManagerLocation> preferredLocations, boolean localOnly)
    {
        // check if there is anything at all in this group assignment
        if (allSlots.isEmpty()) { //如果没有slots,返回
            return null;
        }

        // get the available slots for the group
        Map<ResourceID, List<SharedSlot>> slotsForGroup = availableSlotsPerJid.get(groupId); //取出JobVertex所对应的结构slotsForGroup
        
        if (slotsForGroup == null) { //初始化slotsForGroup
            // we have a new group, so all slots are available
            slotsForGroup = new LinkedHashMap<>();
            availableSlotsPerJid.put(groupId, slotsForGroup);

            for (SharedSlot availableSlot : allSlots) { //因为allSlots是共享的,所以都可以加到slotsForGroup作为可用slots
                putIntoMultiMap(slotsForGroup, availableSlot.getTaskManagerID(), availableSlot); //将availableSlot注册到slotsForGroup,也就是放到其location所对应的list里面
            }
        }
        else if (slotsForGroup.isEmpty()) { //如果slotsForGroup存在,但是没有可用slots
            // the group exists, but nothing is available for that group
            return null;
        }

        // check whether we can schedule the task to a preferred location
        boolean didNotGetPreferred = false;

        if (preferredLocations != null) { //如果有perferred location
            for (TaskManagerLocation location : preferredLocations) { //对每一个具体的prefer location

                // set the flag that we failed a preferred location. If one will be found,
                // we return early anyways and skip the flag evaluation
                didNotGetPreferred = true; //tricky逻辑,如果下面return,这里的设置也没用;如果没返回,说明没有找到prefer的,所以设为true没有问题

                SharedSlot slot = removeFromMultiMap(slotsForGroup, location.getResourceID()); //如果可以在slotsForGroup找到对应prefer location上的slot,为何要remove,因为一个jobvertex不可能有两个task跑在同一个slot上
                if (slot != null && slot.isAlive()) {
                    return new Tuple2<>(slot, Locality.LOCAL); //返回,并且满足prefer,所以是local,local的含义是和prefer在同一个taskmanager上
                }
            }
        }

        // if we want only local assignments, exit now with a "not found" result
        if (didNotGetPreferred && localOnly) { //如果没有找到prefer local,并且需要localonly,返回null
            return null;
        }

        Locality locality = didNotGetPreferred ? Locality.NON_LOCAL : Locality.UNCONSTRAINED; //走到这里,并didNotGetPreferred = false,说明preferredLocations = null,即UNCONSTRAINED,没有约束条件

        // schedule the task to any available location
        SharedSlot slot;
        while ((slot = pollFromMultiMap(slotsForGroup)) != null) { //在不指定taskmanager location的情况下,随意找一个slot
            if (slot.isAlive()) {
                return new Tuple2<>(slot, locality);
            }
        }
        
        // nothing available after all, all slots were dead
        return null;
    }


SharedSlot.allocateSubSlot

见上

 

那么自然有个疑问,allSlots里面的slot哪边来的?

 

addSharedSlotAndAllocateSubSlot

    private SimpleSlot addSharedSlotAndAllocateSubSlot(
            SharedSlot sharedSlot, Locality locality, JobVertexID groupId, CoLocationConstraint constraint) {

        final ResourceID location = sharedSlot.getTaskManagerID();

        synchronized (lock) {
            
            SimpleSlot subSlot;
            AbstractID groupIdForMap;
            
            // add to the total bookkeeping
            if (!allSlots.add(sharedSlot)) { //加到allSlots中
                throw new IllegalArgumentException("Slot was already contained in the assignment group");
            }
                    
            if (constraint == null) {
                // allocate us a sub slot to return
                subSlot = sharedSlot.allocateSubSlot(groupId); //简单的allocate一个simpleSlot
                groupIdForMap = groupId;
            }
            else { //如果有CoLocationConstraint
                
            }
            
            if (subSlot != null) {
                // preserve the locality information
                subSlot.setLocality(locality);
                
                // let the other groups know that this slot exists and that they
                // can place a task into this slot.
                boolean entryForNewJidExists = false;
                
                for (Map.Entry<AbstractID, Map<ResourceID, List<SharedSlot>>> entry : availableSlotsPerJid.entrySet()) {
                    // there is already an entry for this groupID
                    if (entry.getKey().equals(groupIdForMap)) {
                        entryForNewJidExists = true;
                        continue;
                    }

                    Map<ResourceID, List<SharedSlot>> available = entry.getValue();
                    putIntoMultiMap(available, location, sharedSlot); //对于其他的jobVertex,把sharedSlot加上去
                }

                // make sure an empty entry exists for this group, if no other entry exists
                if (!entryForNewJidExists) { //如果存在参数中的groupId,那么就把它的slot信息清空
                    availableSlotsPerJid.put(groupIdForMap, new LinkedHashMap<ResourceID, List<SharedSlot>>());
                }

                return subSlot;
            }
        }
        // end synchronized (lock)
    }

 

而addSharedSlotAndAllocateSubSlot在Scheduler中被调用,

    protected SimpleSlot getNewSlotForSharingGroup(ExecutionVertex vertex,
                                                    Iterable<TaskManagerLocation> requestedLocations,
                                                    SlotSharingGroupAssignment groupAssignment,
                                                    CoLocationConstraint constraint,
                                                    boolean localOnly)
    {
        // we need potentially to loop multiple times, because there may be false positives
        // in the set-with-available-instances
        while (true) {
            Pair<Instance, Locality> instanceLocalityPair = findInstance(requestedLocations, localOnly); //根据locations信息找到local的instance
            
            if (instanceLocalityPair == null) { //如果没有可用的instance,返回null
                // nothing is available
                return null;
            }

            final Instance instanceToUse = instanceLocalityPair.getLeft();
            final Locality locality = instanceLocalityPair.getRight();

            try {
                JobVertexID groupID = vertex.getJobvertexId();
                
                // allocate a shared slot from the instance
                SharedSlot sharedSlot = instanceToUse.allocateSharedSlot(vertex.getJobId(), groupAssignment); //从instance申请一个SharedSlot

                // if the instance has further available slots, re-add it to the set of available resources.
                if (instanceToUse.hasResourcesAvailable()) { //如果这个instance还有多余的资源,再加入instancesWithAvailableResources,下次还能继续用来分配
                    this.instancesWithAvailableResources.put(instanceToUse.getTaskManagerID(), instanceToUse);
                }

                if (sharedSlot != null) {
                    // add the shared slot to the assignment group and allocate a sub-slot
                    SimpleSlot slot = constraint == null ?
                            groupAssignment.addSharedSlotAndAllocateSubSlot(sharedSlot, locality, groupID) : //把分配的SharedSlot加到SlotSharingGroup的SlotSharingGroupAssignment中
                            groupAssignment.addSharedSlotAndAllocateSubSlot(sharedSlot, locality, constraint);

                    if (slot != null) {
                        return slot;
                    }
                    else {
                        // could not add and allocate the sub-slot, so release shared slot
                        sharedSlot.releaseSlot();
                    }
                }
            }
            catch (InstanceDiedException e) {
                // the instance died it has not yet been propagated to this scheduler
                // remove the instance from the set of available instances
                removeInstance(instanceToUse);
            }

            // if we failed to get a slot, fall through the loop
        }
    }

 

getNewSlotForSharingGroup是在当SlotSharingGroup没有可用的slot时,会被调用从instance中分配SharedSlot

 

参考,http://wuchong.me/blog/2016/05/09/flink-internals-understanding-execution-resources/

posted on 2017-04-13 12:17  fxjwind  阅读(6360)  评论(0编辑  收藏  举报