[MPICH] Status of MPICH2 wrt collective communications

[Rajeev Thakur]

The paper you refer to describes most of the recent stuff. As far as I
remember, there have only been minor updates since then (MPICH2 1.0, not
0.971). In any case, upgrading to the latest MPICH2 should give you better
performance than MPICH 1.2.6 because even the basic communication is faster.
In MPICH2, you can see the cutoff points in src/include/mpiimpl.h under
MPIR_BCAST_SHORT_MSG etc. In MPICH-1, they are at the top of
src/coll/intra_fns_new.c

Rajeev

 

> -----Original Message-----
> From: owner-mpich-discuss at mcs.anl.gov 
> [mailto:owner-mpich-discuss at mcs.anl.gov] On Behalf Of Martin 
> Schwinzerl
> Sent: Sunday, April 15, 2007 10:59 AM
> To: mpich-discuss at mcs.anl.gov
> Subject: [MPICH] Status of MPICH2 wrt collective communications
> 
> Dear all,
> 
> I'm searching for information about the used algorithms (esp. 
> the cutoff 
> - points, where
> switching between these algorithms occur) for MPICH2 and 
> would be very 
> grateful
> for any assistence. The most recent piece of information 
> regarding this 
> topic I am aware
> of is the paper by R. Thakur, R. Rabenseifner and W. Gropp [1], which 
> describes the
> implementation for the 0.971 realease of MPICH2 and the 1.2.6 
> version of 
> MPICH1. After
> skimming through the release notes and change logs since 
> these versions,
> my main question is, whether there have been any changes / 
> improvements 
> in this field
> on the algorithmic level since those releases and if so, whether they 
> got publicated.
> 
> I'm interested in this topic due to a couple of reasons (personal 
> interest, discussion with
> our cluster admin about upgrade from MPICH 1.2.6 to MPICH2, 
> discussion 
> with other users
> of our cluster about upgrading from MPICH 1.2.6 to MPICH2, 
> optimising my 
> own parallel code, ... ).
> I just wanted to ask if there is some easier way to get to this 
> information than studying the source
> code directly (which I'm prepared to do, nevertheless :-) ).
> 
> 
> Please accept my appologies, if this is the wrong list to ask such a 
> question.
> 
> Thanks in advance & best regards
> 
> Martin Schwinzerl
> 
> 
> 
> [1] Optimization of Collective Communication Operations in MPICH
> Rajeev Thakur, Rolf Rabenseifner and William Gropp
> Retrieved from 
> http://www-unix.mcs.anl.gov/~thakur/papers/ijhpca-coll.pdf
> 
> 
> 
>