Distributed data structures

A typical computation with the Parallel BGL involves several distributed data structures working in concern. For example, a simple breadth-first search involves the distributed graph data structure containing the graph itself, a distributed queue that manages the traversal through the graph, and a distributed property map that tracks which vertices have already been visited as part of the search.

The Parallel BGL manages these distributed data structures by allowing each of the data structures to attach themselves to the process group itself. When a distributed data structure attaches to the process group, it receives its own copy of the process group that allows the distributed data structure to communicate without colliding with the communications from other distributed data structures. When the process group is synchronized, all of the distributed data structures attached to that process group are automatically synchronized, so that all of the distributed data structures in a computation remain synchronized.

A distributed data structure attaches itself to the process group by creating a copy of the process group and passing an attach_distributed_object flag to the process group constructor. So long as this copy of the process group persists, the distributed data structure is attached the process group. For this reason, most distributed data structures keep a copy of the process group as member data, constructing the member with attach_distributed_object, e.g.,

template<typename ProcessGroup>
struct distributed_data_structure
{
  explicit distributed_data_structure(const ProcessGroup& pg)
    : process_group(pg, boost::parallel::attach_distributed_object())
  { }

private:
  ProcessGroup process_group;
};

Asynchronous receives

Distributed data structures in the Parallel BGL can "asynchronously" receive and process messages before the end of a BSP superstep. Messages can be received any time that a process is inside the process group operations, and the scheduling of message receives is entirely managed by the process group.

Distributed data structures receive messages through "triggers". Triggers are function objects responsible for processing a received message. Each trigger is registered with the trigger method of the process group using a specific message tag (an integer) and the type of data that is expected to be contained within that message. Whenever a message with that tag becomes available, the progress group will call the trigger with the source of the message, the message tag, the data contained in the message, and the "context" of the message.

The majority of triggers have no return value, although it is common that the triggers send messages back to the source process. In certain cases where the trigger's purpose is to immediately reply with a value, the trigger should be registered with the trigger_with_reply method and should return the value that will be sent back to the caller. The trigger_with_reply facility is only useful in conjunction with out-of-band messaging, discussed next.

Out-of-band messaging

The majority of messages sent by the Parallel BGL are sent through the normal send operations, to be received in the next superstep or, in some cases, received "early" by a trigger. These messages are not time-sensitive, so they will be delivered whenever the process group processes them.

Some messages, however, require immediate responses. For example, if a process needs to determine the current value associated with a vertex owned by another process, the first process must send a request to the second process and block while waiting for a response. For such messages, the Parallel BGL's process groups provide an out-of-band messaging mechanism. Out-of-band messages are transmitted immediately, with a much higher priority than other messages. The sending of out-of-band messages can be coupled with a receive operation that waits until the remote process has received the message and sent its reply. For example, in the following code the process sends a message containing the string name to process owner with tag msg_get_descriptor_by_name via an out-of-band message. The receiver of that message will immediately deliver the message via a trigger, that returns the resulting value--a vertex_descriptor--that will be passed back to the process that initiated the communication. The full communication happens immediately, within the current superstep.

std::string name;
vertex_descriptor descriptor;
send_oob_with_reply(process_group, owner, msg_get_descriptor_by_name,
                    name, descriptor);

Process group constructors

ProcessGroup();

Constructs a new process group with a different communication space from any other process group.

ProcessGroup(const ProcessGroup& pg, boost::parallel::attach_distributed_object);

Attaches a new distributed data structure to the process group pg. The resulting process group can be used for communication within that new distributed data structure. When the newly-constructed process group is eventually destroyed, the distributed data structure is detached from the process group.

Triggers

template<typename Type, typename Handler>
  void trigger(int tag, const Handler& handler);

Registers a trigger with the given process group. The trigger will watch for messages with the given tag. When such a message is available, it will be received into a value of type Type, and the function object handler will be invoked with four parameters:

source

The rank of the source process (an int)

tag

The tag used to send the message (also an int)

data:

The data transmitted with the message. The data will have the type specified when the trigger was registered.

context:

The context in which the trigger is executed. This will be a value of type trigger_receive_context, which stages whether the trigger is being executed during synchronization, asynchronously in response to an "early" receive (often to free up communication buffers), or in response to an "out-of-band" message.

Triggers can only be registered by process groups that result from attaching a distributed data structure. A trigger can be invoked in response to either a normal send operation or an out-of-band send operation. There is also a simple trigger interface for defining triggers in common cases.

template<typename Type, typename Handler>
  void trigger_with_reply(int tag, const Handler& handler);

Like the trigger method, registers a trigger with the given process group. The trigger will watch for messages with the given tag. When such a message is available, it will be received into a value of type Type and handler will be invoked, just as with a normal trigger. However, a trigger registered with trigger_with_reply must return a value, which will be immediately sent back to the process that initiated the send resulting in this trigger. Thus, trigger_with_reply should only be used for messages that need immediate responses. These triggers can only be invoked via the out-of-band sends that wait for the reply, via send_oob_with_reply. There is also a simple trigger interface for defining triggers in common cases.

trigger_receive_context trigger_context() const;

Retrieves the current context of the process group with respect to the invocation of triggers. When trc_none, the process group is not currently invoking any triggers. Otherwise, this value describes in what context the currently executing trigger is being invoked.

Helper operations

void poll();

Permits the process group to receive any incomining messages, processing them via triggers. If you have a long-running computation that does not invoke any of the process group's communication routines, you should call poll occasionally to along incoming messages to be processed.

ProcessGroup base() const;

Retrieves the "base" process group for this process group, which is a copy of the underlying process group that does not reference any specific distributed data structure.

Process query

int process_id(const ProcessGroup& pg);

Retrieves the ID (or "rank") of the calling process within the process group. Process IDs are values in the range [0, num_processes(pg)) that uniquely identify the process. Process IDs can be used to initiate communication with another process.

int num_processes(const ProcessGroup& pg);

Returns the number of processes within the process group.

Message transmission

template<typename T>
  void send(const ProcessGroup& pg, int dest, int tag, const T& value);

Sends a message with the given tag and carrying the given value to the process with ID dest in the given process group. All message sends are non-blocking, meaning that this send operation will not block while waiting for the communication to complete. There is no guarantee when the message will be received, except that it will become available to the destination process by the end of the superstep, in the collective call to synchronize.

Any type of value can be transmitted via send, so long as it provides the appropriate functionality to be serialized with the Boost.Serialization library.

template<typename T>
  void receive(const ProcessGroup& pg, int source, int tag, T& value);

Receives a message with the given tag sent from the process source, updating value with the payload of the message. This receive operation can only receive messages sent within the previous superstep via the send operation. If no such message is available at the time receive is called, the program is ill-formed.

optional<std::pair<int, int> > probe(const ProcessGroup& pg);

Determines whether a message is available. The probe operation checks for any messages that were sent in the previous superstep but have not yet been received. If such a message exists, probe returns a (source, tag) pair describing the message. Otherwise, probe will return an empty boost::optional.

A typical use of probe is to continually probe for messages at the beginning of the superstep, receiving and processing those messages until no messages remain.

Synchronization

void synchronize(const ProcessGroup& pg);

The synchronize function is a collective operation that must be invoked by all of the processes within the process group. A call to synchronize marks the end of a superstep in the parallel computation. All messages sent before the end of the superstep will be received into message buffers, and can be processed by the program in the next superstep. None of the processes will leave the synchronize function until all of the processes have entered the function and exchanged messages, so that all processes are always on the same superstep.

Out-of-band communication

template<typename T>
  void send_oob(const ProcessGroup& pg, int dest, int tag, const T& value);

Sends and out-of-band message. This out-of-band send operation acts like the normal send operation, except that out-of-band messages are delivered immediately through a high-priority channel.

template<typename T, typename U>
  void
  send_oob_with_reply(const ProcessGroup& pg, int dest, int
                      tag, const T& send_value, U& receive_value);

Sends an out-of-band message and waits for a reply. The send_oob_with_reply function can only be invoked with message tags that correspond to triggers registered with trigger_with_reply. This operation will send the message immediately (through the high-priority, out-of-band channel), then wait until the remote process sends a reply. The data from the reply is stored into receive_value.

template<typename T>
  void receive_oob(const ProcessGroup& pg, int source, int tag, T& value);

Receives an out-of-band message with the given source and tag. As with the normal receive operation, it is an error to call receive_oob if no message matching the source and tag is available. This routine is used only rarely; for most circumstances, use send_oob_with_reply to perform an immediate send with a reply.

Copyright (C) 2007 Douglas Gregor

Copyright (C) 2007 Matthias Troyer