Information Technology Reference
In-Depth Information
called blocking communication calls, meaning that the message is safely sent
away or received upon the return of the calls.
To enable computation and communication overlap in the present example,
we need to divide the local inner points into two sets. More specifically, the
outermost layer of the local inner points on each subdomain should be com-
puted first. Then, communication for transferring these new values between the
subdomains can be initiated by the non-blocking
MPI Isend
and
MPI Irecv
commands. Thereafter, computation on the remaining local inner points can be
carried out at the same time as communication is being handled. Finally, the
blocking
MPI Wait
command can be used to ensure that all the needed incoming
messages have been received.
Construct a new MPI implementation based the above strategy of overlapping
computation and communication.
