Newsgroups: comp.parallel
From: "Rudnei D. da Cunha" <rudnei@mat.ufrgs.br>
Subject: PVMFMCAST in IBM PVMe 1.2 on an SP/1 using High-Performance-Switch
Organization: Postgraduate Programme on Applied Maths, UFRGS, Brazil
Date: 5 Oct 1995 14:22:10 GMT
Message-ID: <450pmi$rbk@usenet.srv.cis.pitt.edu>

Dear colleagues,

I would appreciate if somebody who has used the IBM SP/1 and its
High-Performance-Switch (HPS) could give me some info on this:

I have an algorithm to compute the outer-product of two vectors
which has been code using IBM PVMe 1.2. At the beginning, each
processor issues a call to PVMFMCAST to the other processors
being used.

Now, I know that in the past the public-domain version of PVMFMCAST
used to be implemented as a series of PVM send calls, instead of
being implemented as (binary-)tree-like scattering of data.

Does anybody know how is it implemented on PVMe? 

As a related question, the SP/1 offers point-to-point communication
between processors through its HPS, which is like an Omega network.
However, as the number of processors grow, in the case of my
specific algorithm, there is certainly some congestion in
the network. My question then is: how long does it take to service
$p$ multicasts in the case when all multicasts are issued (almost)
simultaneously?

Any pointers to published material are welcome.

Please reply by e-mail if possible.
-- 
Rudnei Dias da Cunha
Mathematics Institute & National Supercomputing Centre 
Universidade Federal do Rio Grande do Sul
Av. Bento Goncalves 9500 -- Porto Alegre -- RS -- 91509-900 BRASIL

