Newsgroups: comp.parallel.mpi
From: James Cownie <jcownie@bbn.com>
Subject: Re: reductions deterministic?
Organization: BBN
Date: 19 Mar 1996 09:21:04 GMT
Mime-Version: 1.0
Content-Type: text/plain; charset=iso-8859-1
Content-Transfer-Encoding: 8bit
Message-ID: <4ilua0$9nl@newsgate.dircon.co.uk>

rbarrett@ranger.lanl.gov (Richard Barrett) wrote:
>Bill,
>
>Thanks for the information. However, that's quite an important caveat: getting the
>same answer with a varying number of processors is worth the loss of performance
>for some users. If fact if they don't get it, they will claim our code is "broken" 
>and won't use it. Would it be possible to provide a fully deterministic version 
>option?
>
>Richard
In reply to Bill Gropp's comment
>|> Yes.  But note that deterministic applies to the exact same situation (number
>|> of processors, order of members in group).  It says nothing about differing
>|> numbers of processors (which some would also like, but which seriously 
>|> impacts performance).
>|> 
>|> Bill
>

I'm sorry, but I'm having a bit of trouble understanding what is being
requested.What does it for two different reductions to be deterministic ?
How can MPI possibly ensure that summing (to take an interesting case) two
sets of different sizes can be guaranteed to give the same answer !?
The summation being deterministic means that the additions are always performed
in the same order. So, for instance, sum(1,2,3,4) is always performed as
(1+2) + (3+4) this doesn't imply that it will equal sum(4,1,2,3) (which will
be performed *in the same order* (4+1)+(2+3)). And it certainly says nothing
about how sum(1,1,2,3,3) is performed, or whether that will give the same
result. (Of course you need to read suitable float values for every number
above
to make this interesting. I doubt any implementations will get the exact
examples given wrong !)

-- Jim
James Cownie
BBN


