home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!UB.com!pacbell.com!sgiblab!munnari.oz.au!manuel.anu.edu.au!nimbus!tridge
- From: tridge@nimbus.anu.edu.au (Andrew Tridgell)
- Newsgroups: comp.sys.super
- Subject: Re: LIST OF SITES - NEW RATINGS
- Date: 27 Jan 1993 09:56:26 GMT
- Organization: Australian National University, Canberra
- Lines: 35
- Distribution: inet
- Message-ID: <1k5m8bINNs2@manuel.anu.edu.au>
- References: <1jnktgINNn60@uniwa.uwa.edu.au> <1993Jan22.153944.5686@chpc.utexas.edu>
- NNTP-Posting-Host: 150.203.15.21
-
- >Since many of those sites included [not very interesting] peak MFLOPS
- >ratings for Intel Paragon and TMC CM-5 machines, but Dongarra's latest
- >[January 13, 1993] Linpack report STILL has no Linpack numbers for the
- >Paragons and CM-5s (even in Table 3, where one can make the matrices
- >as large as they wish!), would some owner of these machines please
- >give us the Linpack results for those two systems?
-
- I haven't tried linpack itself but I have coded some vector and matrix
- routines. On a CM-5 with vector units I have achieved 75MFlops per node
- for vector*matrix, matrix*vector and matrix*matrix routines. I have
- achieved 45MFlops for an outer product of two vectors. These are all for
- double precision numbers. I believe the quoted peak performance is
- 128MFlops per node. I am not a experienced CM5 programmer (these were my
- first vector unit routines) and expect that much better results could be
- obtained by someone with more experience. In particular I suspect that
- the pipelining of instructions could give me a big improvement.
-
- I coded these routines in CDPEAC and tested with vector lengths of
- around 100 and matrices of approx 100*100.
-
- On a related issue - has anyone got any code to copy data from scalar
- memory to the DP stack or heap on the CM5 at better than 16Mb per
- second? I really need to copy data much faster than that. This is
- really only an issue if you are programming using the CMMD message
- passing library.
-
-
- Andrew
-
-
- --
- =-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
- Andrew Tridgell CSLab, Research School of Physical Sciences
- Andrew.Tridgell@anu.edu.au Australian National University (x3064)
- =-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
-