Comparing: Collective communications

Collective Communication in PVM

Dynamic process group functions are a new feature available in PVM3

Any PVM task can join or leave a group without having to inform other tasks.

Joining and leaving groups in PVM

pvmfjoingroup(group, inum)
Where:
group is a character string of an existing group.
inum is the instance number of this task joining the group.
pvmflvgroup(group, info)
Where:
info is an integer status code.

Broadcasting messages in PVM

Tasks can broadcast messages to groups whether or not they belong to that group.

pvmfbcast(group, msgtag, info)
This is an asynchronous call and computation on the sending process resumes as soon as the message is sent.

Collective communication in MPI

Collective communication in MPI involves every process in the specified communicator.

To perform collective communication on a subset of processes a new communicator has to be created.

Distribution of data amongst processes in MPI

There exists a number of routines to distribute data amongst processes; these include:

MPI_BCAST
MPI_SCATTER
MPI_GATHER
MPI_ALLGATHER

MPI reduction operators

Global Reduction Operators.
These are used to compute a result that involves data distributed across a group of processes.
Reduction operators may be pre-defined or user-defined.

Predefined operators include:

	MPI_MAX 	Maximum
	MPI_MIN 	Minimum
	MPI_SUM 	Sum
	MPI_PROD	Product
	MPI_IAND	Logical AND
	MPI_BAND	Bitwise AND

Submitted by Mark Johnston,
last updated on 10 December 1994.