On 2002-09-03 Jeff Squyres wrote:
[N is the number of processors.]
> If N gets to be large, you may want to do something more clever than a
> straight serial send. For example, you may want to use MPI_Scatter()
> to do the sending, which, even though you'll have the send the
> "master"'s bytes to itself, can be faster if the MPI implementation
> uses a tree-based sending algorithm, rather than a linear algorithm
> (which LAM does, if N > 4). That gets a little more complicated to
> calculate, so I'll leave that as an exercise for the reader (and this
> mail is long enough already!).
> :-)
The context of this message was: On 2002-08-30 Robbie Nadig wrote:
> I have a large file that I want to search for specific patterns. The
> results of the search are presented to the user in a TCL GUI.
>
> Currently, searching for patterns takes a long time and hence makes
> interactive use nearly impossible.
>
> I was wondering if it would be easy to split the large file and load the
> file into programs on a LAM-mpi node, that way each node performs the
> search on a small segment of the entire file.
My question is: how will a tree-based sending algorithm (within
MPI_Scatter()) save any time? Isn't the limiting factor the number of
bytes transmitted from the master, which will always be the same?
Thanks,
Paul Howard
Microway, Inc.
_______________________________________________
This list is archived at http://www.lam-mpi.org/MailArchives/lam/
|
