org.apache.giraph.comm.messages.queue

Class AsyncMessageStoreWrapper<I extends org.apache.hadoop.io.WritableComparable,M extends org.apache.hadoop.io.Writable>

java.lang.Object

org.apache.giraph.comm.messages.queue.AsyncMessageStoreWrapper<I,M>

Type Parameters:

I - Vertex id

M - Message data

All Implemented Interfaces:

MessageStore<I,M>

public final class AsyncMessageStoreWrapper<I extends org.apache.hadoop.io.WritableComparable,M extends org.apache.hadoop.io.Writable>
extends Object
implements MessageStore<I,M>

This class decouples message receiving and processing into separate threads thus reducing contention. It does not provide message store functionality itself, rather providing a wrapper around existing message stores that can now be used in async mode with only slight modifications.

Constructor Summary

Constructors

Constructor and Description
AsyncMessageStoreWrapper(MessageStore<I,M> store, Iterable<Integer> partitions, int threadCount) Constructs async wrapper around existing message store object.

Method Summary

Modifier and Type	Method and Description
void	addMessage(I vertexId, M message) Adds a message for a particular vertex The method is used by InternalMessageStore to send local messages; for the general case, use a more efficient addPartitionMessages
void	addPartitionMessages(int partitionId, VertexIdMessages<I,M> messages) Adds messages for partition
void	clearAll() Clears all resources used by this store.
void	clearPartition(int partitionId) Clears messages for a partition.
void	clearVertexMessages(I vertexId) Clears messages for a vertex.
void	finalizeStore() Called before start of computation in bspworker Since it is run from a single thread while the store is not being accessed by any other thread - this is ensured to be thread-safe
Iterable<I>	getPartitionDestinationVertices(int partitionId) Gets vertex ids from selected partition which we have messages for
Iterable<M>	getVertexMessages(I vertexId) Gets messages for a vertex.
boolean	hasMessagesForPartition(int partitionId) Check if we have messages for some partition
boolean	hasMessagesForVertex(I vertexId) Check if we have messages for some vertex
boolean	isPointerListEncoding() True if this message-store encodes messages as a list of long pointers to compact serialized messages
void	readFieldsForPartition(DataInput in, int partitionId) Deserialize messages for one partition
void	waitToComplete() Wait till all messages are processed and all queues are empty.
void	writePartition(DataOutput out, int partitionId) Serialize messages for one partition.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Detail

AsyncMessageStoreWrapper

public AsyncMessageStoreWrapper(MessageStore<I,M> store,
                                Iterable<Integer> partitions,
                                int threadCount)

Constructs async wrapper around existing message store object. Requires partition list and number of threads to properly initialize background threads and assign partitions. Partitions are assigned to threads in round-robin fashion. It guarantees that all threads have almost the same number of partitions (+-1) no matter how partitions are assigned to this worker.

Parameters:

store - underlying message store to be used in computation

partitions - partitions assigned to this worker

threadCount - number of threads that will be used to process messages.

Method Detail

isPointerListEncoding

public boolean isPointerListEncoding()

Description copied from interface: MessageStore

True if this message-store encodes messages as a list of long pointers to compact serialized messages

Specified by:

isPointerListEncoding in interface MessageStore<I extends org.apache.hadoop.io.WritableComparable,M extends org.apache.hadoop.io.Writable>

Returns:

true if we encode messages as a list of pointers

getVertexMessages

public Iterable<M> getVertexMessages(I vertexId)

Description copied from interface: MessageStore

Gets messages for a vertex. The lifetime of every message is only guaranteed until the iterator's next() method is called. Do not hold references to objects returned by this iterator.

Specified by:

getVertexMessages in interface MessageStore<I extends org.apache.hadoop.io.WritableComparable,M extends org.apache.hadoop.io.Writable>

Parameters:

vertexId - Vertex id for which we want to get messages

Returns:

Iterable of messages for a vertex id

clearVertexMessages

public void clearVertexMessages(I vertexId)

Description copied from interface: MessageStore

Clears messages for a vertex.

Specified by:

clearVertexMessages in interface MessageStore<I extends org.apache.hadoop.io.WritableComparable,M extends org.apache.hadoop.io.Writable>

Parameters:

vertexId - Vertex id for which we want to clear messages

clearAll

public void clearAll()

Description copied from interface: MessageStore

Clears all resources used by this store.

Specified by:

clearAll in interface MessageStore<I extends org.apache.hadoop.io.WritableComparable,M extends org.apache.hadoop.io.Writable>

hasMessagesForVertex

public boolean hasMessagesForVertex(I vertexId)

Description copied from interface: MessageStore

Check if we have messages for some vertex

Specified by:

hasMessagesForVertex in interface MessageStore<I extends org.apache.hadoop.io.WritableComparable,M extends org.apache.hadoop.io.Writable>

Parameters:

vertexId - Id of vertex which we want to check

Returns:

True iff we have messages for vertex with required id

hasMessagesForPartition

public boolean hasMessagesForPartition(int partitionId)

Description copied from interface: MessageStore

Check if we have messages for some partition

Specified by:

hasMessagesForPartition in interface MessageStore<I extends org.apache.hadoop.io.WritableComparable,M extends org.apache.hadoop.io.Writable>

Parameters:

partitionId - Id of partition which we want to check

Returns:

True iff we have messages for the given partition

addPartitionMessages

public void addPartitionMessages(int partitionId,
                                 VertexIdMessages<I,M> messages)

Description copied from interface: MessageStore

Adds messages for partition

Specified by:

addPartitionMessages in interface MessageStore<I extends org.apache.hadoop.io.WritableComparable,M extends org.apache.hadoop.io.Writable>

Parameters:

partitionId - Id of partition

messages - Collection of vertex ids and messages we want to add

addMessage

public void addMessage(I vertexId,
                       M message)
                throws IOException

Description copied from interface: MessageStore

Adds a message for a particular vertex The method is used by InternalMessageStore to send local messages; for the general case, use a more efficient addPartitionMessages

Specified by:

addMessage in interface MessageStore<I extends org.apache.hadoop.io.WritableComparable,M extends org.apache.hadoop.io.Writable>

Parameters:

vertexId - Id of target vertex

message - A message to send

Throws:

IOException

finalizeStore

public void finalizeStore()

Description copied from interface: MessageStore

Called before start of computation in bspworker Since it is run from a single thread while the store is not being accessed by any other thread - this is ensured to be thread-safe

Specified by:

finalizeStore in interface MessageStore<I extends org.apache.hadoop.io.WritableComparable,M extends org.apache.hadoop.io.Writable>

getPartitionDestinationVertices

public Iterable<I> getPartitionDestinationVertices(int partitionId)

Description copied from interface: MessageStore

Gets vertex ids from selected partition which we have messages for

Specified by:

getPartitionDestinationVertices in interface MessageStore<I extends org.apache.hadoop.io.WritableComparable,M extends org.apache.hadoop.io.Writable>

Parameters:

partitionId - Id of partition

Returns:

Iterable over vertex ids which we have messages for

clearPartition

public void clearPartition(int partitionId)

Description copied from interface: MessageStore

Clears messages for a partition.

Specified by:

clearPartition in interface MessageStore<I extends org.apache.hadoop.io.WritableComparable,M extends org.apache.hadoop.io.Writable>

Parameters:

partitionId - Partition id for which we want to clear messages

writePartition

public void writePartition(DataOutput out,
                           int partitionId)
                    throws IOException

Description copied from interface: MessageStore

Serialize messages for one partition.

Specified by:

writePartition in interface MessageStore<I extends org.apache.hadoop.io.WritableComparable,M extends org.apache.hadoop.io.Writable>

Parameters:

out - DataOutput to serialize this object into

partitionId - Id of partition

Throws:

IOException

readFieldsForPartition

public void readFieldsForPartition(DataInput in,
                                   int partitionId)
                            throws IOException

Description copied from interface: MessageStore

Deserialize messages for one partition

Specified by:

readFieldsForPartition in interface MessageStore<I extends org.apache.hadoop.io.WritableComparable,M extends org.apache.hadoop.io.Writable>

Parameters:

in - DataInput to deserialize this object from.

partitionId - Id of partition

Throws:

IOException

waitToComplete

public void waitToComplete()

Wait till all messages are processed and all queues are empty.