Re: [bgl-discuss] MPI library on BG/L

[Sayaka Akioka <tobita@xxxxxxxxxxx>]

Hi Sameer,

I am so glad with your help!
Yes, actually, I tried TAU (old version) for the first trial.
The reason why I was forced to change mpich itself is that
the person who will use that profile data cares so much about
the cost of the wrappers.

I myself think the cost is not so expensive, but he does.
Could you show some concrete (but simple one is enough) data
how this wrapper costs? (And I know the cost is a little
enough to ignore :-) )
In that case, I can try to persuade him to use TAU profiling
again, with your help.

I personally think TAU is a nice tool for profiling and makes
things easy. If I can persuade him, I would like to use TAU.

Cheers,
Sayaka


Sameer Shende wrote:
> Hi Sayaka,
> 	Would using the PMPI profiling interface help in this case? You'd
> implement your own collective MPI functions and do what a tool such as TAU
> does - interpose your library between the application and the MPICH
> library. Many profiling and tracing tools use this approach to add timing
> calls to native MPI libraries without changing the implementation of
> MPICH. They internally call the namshifted interface:
>
> #pragma weak MPI_Send = PMPI_Send
>
> int MPI_Send(...)
> { /* original implementation */
> }
>
> wrapper library:
> int MPI_Send(...)
> {
>   timer_start();
>   int ret_value = PMPI_Send(...);
>   timer_stop();
>   return ret_value:
> }
> and while linking in:
> -lWrapperlib -lpmpich -lmpich
>
> 	If you use this approach and don't invoke the corresponding
> PMPI_... call, you'd add your code in the wrapper library. It'll make it
> portable to other MPI implementations as well. What do you feel?
> 	If you are interested in seeing how TAU does this, see
> src/Profile/Makefile.wrapper TauMpi.c TauFMpi.c files.
> 	Thanks,
> 	- Sameer
>
> On Mon, 5 Sep 2005, Sayaka Akioka wrote:
>
>
> > Hello,
> >
> > I need your help. But it might be better to send this mail
> > to bgl-discuss.
> >
> > Now I am trying to run my code which utilizes MPI library.
> > What I would like to do in my code is to use my own collective
> > MPI functions instead of MPI standard collective functions.
> > I have done this work on other PC clusters, but now I am trying
> > to do the same thing on BG/L. I have trouble getting compiled my
> > codes.
> >
> > For example, in MPIC_Send, MPIC_Recv, MPIC_Sendrecv, MPID_Irecv,
> > MPID_Isend, MPID_Send, MPID_Recv, references to functions (their
> > name start with "MPIDI_CH3") can not be found. I used mpicc on
> > front end machine, under /bgl/BlueLight/ppcfloor/bglsys.
> >
> > I am guessing this is because I did something wrong, or MPI
> > library on BG/L should be the special version. To get my codes
> > compiled, what should I do?
> >
> > I am attaching the error messages.
> >
> > Thank you,
> > Sayaka
> >
> > --
> > Sayaka Akioka
> > e-mail: akioka@xxxxxxxxxxx
>
>
> - --------------------------------------------------------------------
> To add or remove yourself from this mailing list, use the 'notifyme'
> command on any BGL machine. To remove: notifyme -n, to add: notifyme -y.


--
Sayaka (Tobita) Akioka
Post-doctoral Research Associate
The Pennsylvania State University
e-mail: tobita@xxxxxxxxxxx

- --------------------------------------------------------------------
To add or remove yourself from this mailing list, use the 'notifyme'
command on any BGL machine. To remove: notifyme -n, to add: notifyme -y.