home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!usc!sdd.hp.com!swrinde!gatech!hubcap!fpst
- From: janowsky@math.rutgers.edu (Steven Janowsky)
- Newsgroups: comp.parallel
- Subject: Re: Could I compete the low-end CM5 with multiple SPARCstations?
- Message-ID: <1992Jul27.140950.26223@hubcap.clemson.edu>
- Date: 24 Jul 92 18:19:47 GMT
- References: <1992Jul23.134928.5014@hubcap.clemson.edu> <1992Jul23.154028.11687@unocal.com>
- Sender: fpst@hubcap.clemson.edu (Steve Stevenson)
- Organization: Rutgers Univ., New Brunswick, N.J.
- Lines: 33
- Approved: parallel@hubcap.clemson.edu
-
-
- In article <1992Jul23.154028.11687@unocal.com> stgprao@xing.unocal.com (Richard Ottolini) writes:
-
- > (1) Each node of CM-5 contains a 128 megaflop array processor (actually four
- > 32 MF vector chips if you buy the full configuration.) This an order of
- > magnitude faster than the fastest SPARC chip.
-
- Current CM5 configurations DO NOT have these vector chips. Or at least
- none of the CM5's that I have access to have them yet.
- Until they arrive...
-
- > (2) The communication channels between the nodes scales with the number of
- > processors on a CM-5 in contrast to a workstation network that subdivides
- > traffic on a single network. We do distributed number crunching at our
- > site on a workstation network. Network saturation often occurs when the
- > number of nodes are in the low double digits.
-
- 32 = low double digits. For certain low-communication applications, the
- network of SS's should be better. For my applications, where I want to
- get STATISTICS, I'd be better running 32 independent copies of the
- program on each machine and then have ZERO communication.
-
- > (3) SOFTWARE, SOFTWARE, SOFTWARE! to paraphrase a real estate maxim.
- > TM is attempting to implement their data-parallel programming languages-
- > TM-FORTRAN-90/HPF and others- that hide need to worry a lot about how to
- > distribute data and communications.
-
- How much overhead do you think the compiler eats up ? I.e. do you
- really get 32 procs of power? On the CM2/200 I am playing compiler
- games in order to approach theoretical max speed. Those more dedicated
- than me make PARIS calls directly.
-
- SJ
-