On Tue 08 May 2012 02:53:28 AM CEST, Jeff Squyres (jsquyres) wrote:
> I *believe* that this value came out of IB hardware history. I.e., when we wrote the openib BTL, some empirical testing showed that 12k was a good value.
Thanks a lot for the clarification. Do you know if this empirical tests
has been published somehow/somewhere?
> On May 7, 2012, at 9:25 AM, Simone Pellegrini wrote:
>> I have one of those 1M dollar questions I guess, but why the eager limit threshold for Infiniband is set to 12KB by default in OpenMPI? I would like to know where this value comes from. I am not wondering whether this is a good setting for this parameter, but just why this is considered to be a good trade-off for all Infiniband clusters. What I would like to know at the end of the story is whether this is an empirical value or if there are architecture characteristics of IB which make this value come out of pen and paper work (any pointer would be appreciated).
>> thanks in advance for the reply.
>> Simone P.
>> users mailing list