Can I change the eager limit in mpiexec

[Ying Chen]

Hi Pete,

Thanks for your help! I have figured out the environment variable for the
eager sends. That is "GMPI_EAGER_SIZE" for mpirun.ch_gm. However, I still
get problem when I use mpi_bcast to send a lot of short messages from root
node to other nodes. It seems that it is good at beginning, but it has
problem later and the values received by the nodes are not correct.

The same code doesn't have such problem for IBM SP with poe, and have no
problem for our beowulf cluster with ethernet. After we updated myrinet
and mpich, we have such problem. Do you have any suggestion about it?

thanks,

Ying

 On Thu, 14 Nov 2002, Pete Wyckoff wrote:

> yingchen at Stanford.EDU said:
> > I am a graduate student at Stanford. We installed mpiexec in our compaq
> > alpha linux cluster. I am testing it with some f90 code. Do you know if
> > there is any method I can increase the eager limit for the synchronization
> > calls. For IBM cluster with poe, it has a option called eager_limit. I
> > want to increase it to 64k if it is possible.
>
> Sorry I'm not sure what you're talking about.  Are you thinking of the
> message size limit which decides between eager and rendez-vous sends
> used by most MPI implementations?
>
> If you can figure out an environment variable, just set it in the shell
> before calling mpiexec and it will be propagated to the applications and
> hence to the MPI library.
>
> 		-- Pete
>