From: Jean-Paul Dzisiak Newsgroups: comp.parallel.mpi Subject: performance of MPICH 1.1 under AIX 4.3 [HELP !] Date: Fri, 26 Feb 1999 16:59:51 +0100 Organization: Environment Center - ULg Message-Id: <36D6C4F7.BE30B4B4@ulg.ac.be> Mime-Version: 1.0 Content-Type: multipart/alternative; boundary="------------7766A07DD74D2D2D953FE03A" Xref: ukc comp.parallel.mpi:4676 --------------7766A07DD74D2D2D953FE03A Content-Type: text/plain; charset=iso-8859-1 Content-Transfer-Encoding: quoted-printable Hello ! I've 3 processes running on an AIX 4.3 IBM workstation. I'm using the "p4" device: [process 1 (P1)] || [process 2 (P2)] <=3D=3D=3D> [process 3 (P3)] All the messages use MPI_SEND / MPI_RECV Per iteration, I have: i) P1 sends P2 600 blocks of 30 reals ii) P2 sends P3 1 block of 300 reals iii) P3 sends P1 1 block of 130 reals When I don't activate messages P3 to P2 [ P2 =3D> P3], my whole enviromen= t run very fast. OK. but when I activate messages P3 to P2 (even a single REAL !) [P2 <=3D> P3], performances collapse Why a single MPI_SEND/MPI_RECV message (that send only 130 real !) significantly slows down the processes ? Is there a known bug ? Do I have to update MPICH to 1.2 ? Help would be very appreciated ... -- Jean-Paul Dzisiak CEME - Universit=E9 de Li=E8ge Centre d'Etude et de Mod=E9lisation de l'Environnement Sart-Tilman B5 B-4000 Li=E8ge BELGIQUE T=E9l +32 (0) 4 366 23 56 Fax +32 (0) 4 366 23 55 e-mail JP.Dzisiak@ulg.ac.be --------------7766A07DD74D2D2D953FE03A Content-Type: text/html; charset=us-ascii Content-Transfer-Encoding: 7bit Hello !

I've 3 processes running on an AIX 4.3 IBM workstation. I'm using the "p4" device:

[process 1 (P1)]
   ||
[process 2 (P2)]  <===> [process 3 (P3)]

All the messages use MPI_SEND / MPI_RECV

Per iteration, I have:
  i)   P1 sends P2 600 blocks of 30 reals
  ii)  P2 sends P3 1 block of 300 reals
  iii) P3 sends P1 1 block of 130 reals

When I don't activate messages P3 to P2 [ P2 => P3], my whole enviroment run very fast. OK.
but when I activate messages P3 to P2 (even a single REAL !) [P2 <=> P3], performances collapse

Why a single MPI_SEND/MPI_RECV message (that send only 130 real !) significantly slows down
the processes ? Is there a known bug ? Do I have to update MPICH to 1.2 ? 

Help would be very appreciated ...
-- 
Jean-Paul Dzisiak

CEME - Université de Liège
Centre d'Etude et de Modélisation de l'Environnement
Sart-Tilman B5
B-4000 Liège
BELGIQUE

Tél +32 (0) 4 366 23 56
Fax +32 (0) 4 366 23 55
e-mail JP.Dzisiak@ulg.ac.be
  --------------7766A07DD74D2D2D953FE03A--