From: Thomas Albers (talbers_at_binghamton.edu)
Date: Tue Jun 14 2011 - 13:04:29 CDT
> In this case pe 7 caught a segfault signal. There might be some useful
> information in the stack trace. Building a debug version would give you
> more information on where the segfault happened.
How is that done?
> Do you see an actual performance improvement from Open-MX?
Linux-x86_64, precompiled, 1 core/machine:
TIMING: 500 CPU: 209.475, 0.336699/step Wall: 211.611, 0.338333/step
Open-MX, gcc 1 core/machine:
TIMING: 500 CPU: 262.249, 0.471274/step Wall: 262.249, 0.471274/step,
Afraid not. I'm surprised to see that the performance penalty is that
high. One would predict that (Open)MX has a lower latency than TCP/IP,
so this result is quite stunning. Could that be the overhead from OpenMPI?
Something else - the notes say that charm++ can be built with Myrinet MX
support. When you do try
./build charm++ net-linux-x86_64 mx -j4 --with-production
this fails with
machine.c: At top level:
machine.c:2425: error: redefinition of 'CmiBarrier'
machine-mx.c:737: note: previous definition of 'CmiBarrier' was here
machine.c:2455: error: redefinition of 'CmiBarrierZero'
machine-mx.c:784: note: previous definition of 'CmiBarrierZero' was here
Fatal Error by charmc in directory
Is this a Charm++ bug?
> Does the simulation you are running scale better with an InfiniBand
I'm afraid that that is something we can't afford.
This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:57:18 CST