I see. I didn't know that. Thanks, Rajeev On Apr 7, 2013, at 10:21 PM, Robert van de Geijn wrote:
Rajeev,
I pointed out to Jack that Fox et al. actually discovered the scatter/allgather broadcast, and that I merely popularized it. So he went and tracked it down.
Robert
On Apr 7, 2013, at 10:19 PM, Rajeev Thakur wrote:
Jack, At the time that code was originally written around 10 years ago, Rabenseifner's algorithm for long-message reduce was not so well known, which is why I think I mentioned it in the code, whereas the long-message broadcast algorithm was relatively well known. A paper describing the collectives in MPICH, published in early 2005, cites two of Robert's papers for the long-message broadcast implemented as scatter-allgather: http://www.mcs.anl.gov/~thakur/papers/ijhpca-coll.pdf.
Regards, Rajeev
On Apr 7, 2013, at 9:48 PM, Jack Poulson wrote:
Dear all,
I am putting together a lecture on collective communication and was perusing the MPICH 3.0.3 source code when I noticed that the only references for the long-message implementations of Reduce and Broadcast seem to be "Rabenseifer's algorithm" for Reduce, with a pointer to the following webpage and corresponding 1997 publication: https://fs.hlrs.de/projects/par/mpi//myreduce.html
This seems to be a bit strange to me, since this is essentially the Scatter composed with Allgather approach to Broadcast implemented in reverse, which was proposed by Fox, Furmanski, and Walker nearly a decade earlier in "Optimal matrix algorithms on homogeneous hypercube" (see Figs. 1 and 2, where the term "expand" is used instead of the modern "Allgather"): http://dl.acm.org/citation.cfm?id=63125
It would seem to me that, if the source code is to include academic references, it would make sense to also cite this (much earlier) paper.
Best Regards, Jack Poulson _______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss