In Open MPI, we detect when the source and receiver side share the
same architecture. Then, the datatype engine avoids type conversion
(let say XDR, ou htonl) and sends raw binary data on the link. If the
architecture mismatch, the datatype engine is forced to use some
conversion which will lead to some overhead. On top of this, you can
expect a slight overhead when using 64bit integer computation compared
to 32bits (this is a general case for any 64bit application, due to
more pressure on the memory bandwidth, though it does not change
anything on floating point performance).
Le 11 avr. 08 à 05:34, jody a écrit :
> What is the cause of this performance penalty?
> On Fri, Apr 11, 2008 at 1:44 AM, Aurélien Bouteiller
> <bouteill_at_[hidden]> wrote:
>> Open MPI can manage heterogeneous system. Though you prefer to avoid
>> this because it has a performance penalty. I suggest you compile on
>> the 32bit machine and use the same version everywhere.
>> Le 10 avr. 08 à 18:09, clarkmpi_at_[hidden] a écrit :
>>> Thanks to those who answered my post in the past. I have to admit
>>> that you lost me about half way through the thread.
>>> I was able to get 2 of my systems cranked up and was about to put a
>>> third system online when I remembered it was running x64 version of
>>> Can I just recompile the code on the x64 system and put it in the
>>> same home directory used by all the systems? I'm not sharing the
>>> directory across systems, but after doing this three or four times
>>> across just 2 systems, I can see why sharing would be advantages.
>>> users mailing list
>> users mailing list
> users mailing list