Hello,
thanks for your response, I just wanted to check that the syntax
of my code was ok, with the mpirun command switches, as the user
documentation says to use another value other than default '-1' for
the 'rpi_tcp_sockbuf', so I set it as follows:
$ mpirun -ssi rpi_tcp_sockbuf 0 N <executable name>
and also then I included the socket setting in the code as they say, I
just
wanted to know if this is the correct way to do it with the mpirun
command.
(in program code...)
sock = socket( PF_INET, SOCK_STREAM, 0 );
ierr_setsock = setsockopt(sock,0, SO_SNDBUF, &val,
131072) ;
ierr_setsock2 = setsockopt(sock,0, SO_RCVBUF, &val,
131072) ;
I am checking performance issues for collective communications as
part of a project and we are investigating the causes of congestion,
so I am looking at where the bottlenecks are and see if the socket size
matters,
thanks for your help.
----- Original Message -----
From: Rajagopal Natarajan <rajagopal.n_at_[hidden]>
Date: Tuesday, July 3, 2007 6:32 am
Subject: Re: [lam-devel] changing tcp socket sizes with rpi_tcp_sockbuf
To: LAM/MPI development issues <lam-devel_at_[hidden]>
> On 6/27/07, Maureen O Flynn <maureen.oflynn_at_[hidden]> wrote:
> >
> > Is this correct? I am not sure it works as my performance figures
> > dont indicate any change so perhaps they are not being resized at
> > all. Any suggestions regarding how to do this would be great.
>
>
> To be sure if it works or not, use some really network intensive
> MPI job
> such as GMX Alltoall
>
http://www.gromacs.org/component/option,com_docman/task,doc_download/gi
d,67/
>
> Its a ping pong test which sends msgs of different sizes to all
> nodes in the
> cluster and generates some high traffic.. varying the packet sizes
and
> testing would help check if the buffer size is really affected.
> may be you
> ran a processor intensive task, and that might be the reason why
> you didnt
> find much difference. Or may be check if in the program that you
> run, the
> push bit(PSH) is enabled.. In that case even if you make your
> buffer large,
> every packet would be sent off immediately, and you cant see the
> effect..
>
> --
> N. Rajagopal,
> Visit me at http://users.kaski-net.net/~raj/
>
|
