Em 30-06-2017 17:28, Brice Goglin escreveu:
Le 30/06/2017 22:08, fabricio a écrit :
Em 30-06-2017 16:21, Brice Goglin escreveu:
Yes, it's possible but very easy. Before we go that way:
Can you also pass HWLOC_COMPONENTS_VERBOSE=1 in the environment and send
the verbose output?
Hello
We have seen _many_ reports like these. But there are different kinds of
errors. As far as I understand:
* Julio's error is caused by the Linux kernel improperly reporting L3
cache affinities. It's specific to multi-socket 12-core processors
because the kernel makes invalid assumptions
Le 30/06/2017 22:08, fabricio a écrit :
> Em 30-06-2017 16:21, Brice Goglin escreveu:
>> Yes, it's possible but very easy. Before we go that way:
>> Can you also pass HWLOC_COMPONENTS_VERBOSE=1 in the environment and send
>> the verbose output?
>
>
We (Georgia Tech) too have been observing this on 16-core AMD AbuDhabi machines
(6378). We weren’t aware of HWLOC_COMPONENTS workaround, which seems to
mitigate the issue.
Before:
# ./lstopo
* hwloc has encountered
Em 29-06-2017 02:24, Brice Goglin escreveu:
Hello Brice
I'm still seeing this error message even when passing the
HWLOC_COMPONENTS=x86 variable.
Is it possible to generate a xml file that can silence this error?
TIA,
Fabricio
___
hwloc-users
Well, yes and no. Yes, your cpu loads will balance better across nodes
(balancing across sockets doesn’t do much for you). However, your overall
application performance may be the poorest in that arrangement if your app uses
a lot of communication as the layout minimizes the use of shared
Hello Ralph,
Thank you for your comments.
My understanding, from reading Jeff's blog on V1.5 processor affinity, is that
the bindings in
Example 1 balance the load better than the bindings in Example 2.
Therefore I would like to obtain the bindings in Example 1, but using Open MPI
2.1.1, and