Re: communication pattern causes corrupted data

On Tue, 2003-03-25 at 18:06, Greg Lindahl wrote:
> On Tue, Mar 25, 2003 at 02:12:44PM -0800, Joshua J. England wrote:
> 
> > I did a full recompile of the redhat 7.1 glibc explicitly setting
> > host=alphaev56-unknown-linux-gnu.  Just to be sure, I also recompiled
> > gm-1.6.3 and mpich-gm-1.2.4..8a.
> > 
> > The problem still occurs.
> 
> Hm. Did you check to see if memcopy and friends actually use byte
> accesses for the ev56 in that old glibc?

You almost certainly have a newer version of glibc that we used.  Our
experience was with RedHat-6.2.  The newer one could very well have
ev4-compatible assembly code for those.  Hmm... I vaguely recall seeing
different assembly versions of the copies, at least in some version of
glibc.  Did you run configure again, or just change the compile options
in the make file?  You may need to reconfigure, to change some links.

> An alternate test would be to pad your buffer variables with an extra
> 8 bytes. If the problem Dave mentions is present, then the corruption
> won't touch the actual data.

This was part of the testing that we did to diagnose it, and is a very
good idea.

> -- greg
> 
> _______________________________________________
> Myrinet mailing list
> [email protected]
> http://email.osc.edu/mailman/listinfo/myrinet
-- 
David B. Ritch
High Performance Technologies, Inc.