[mpich-discuss] Citations for Scatter/Allgather Broadcast and ReduceScatter/Gather Reduce

[Rajeev Thakur]

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     discuss at mpich.org
>>> To manage subscription options or unsubscribe:
>>> https://lists.mpich.org/mailman/listinfo/discuss
>