home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.sys.sgi
- Path: sparky!uunet!zaphod.mps.ohio-state.edu!uwm.edu!ux1.cso.uiuc.edu!news.cso.uiuc.edu!uimrl7.mrl.uiuc.edu!ercolessi
- From: ercolessi@uimrl3.mrl.uiuc.edu (furio ercolessi)
- Subject: Re: Indigo R4k/IBM 6k comparison
- References: <8378@news.duke.edu>
- Message-ID: <C0GMqw.6LF@news.cso.uiuc.edu>
- Sender: usenet@news.cso.uiuc.edu (Net Noise owner)
- Reply-To: ercolessi@uimrl3.mrl.uiuc.edu (furio ercolessi)
- Organization: MRL - UIUC
- Date: Thu, 7 Jan 1993 01:28:06 GMT
- Lines: 70
-
- In article <8378@news.duke.edu>, rla@canctr.mc.duke.edu writes:
- |>Greetings. We are considering a purchase of a set of 4 workstations to be
- |>linked together. Objectives are mainly processing (number-crunching) power
- |>plus the ability to run the GCG DNA sequence analysis suite of programs.
- |>The two remaining candidates are SGI (R4000-based) vs. IBM RS6000s.
- |>
- |>Anyone with experience on both platforms, including service and reliability
- |>and UNIX performance, please let me know your thoughts. Overall system
- |>speed is a major consideration.
-
- I can perhaps say something on number-crunching, since we have run a
- few benchmarks in Trieste. We do heavy computational physics there.
- *The results are very strongly application dependent*, much more than
- we expected.
-
- We have found that the IBMs have a very good performance in linear
- algebra calculations, particularly with big-sized arrays. For example,
- the 560 processor scores 84 Mflops in Linpack 1000x1000, which is
- indeed impressive. HP and SGI are far behind, typically crawling around
- 20 Mflops on this benchmark [these data can be obtained as a PostScript
- file by sending a mail to netlib@ornl.gov, containing the message
- 'send performance from benchmark']. The case of DEC Alpha is
- a bit mysterious because it is reported to be 80 MFlops for the
- 3000-500 machine, but we could get only 11. In the other cases
- we were usually able to reproduce the table results.
- Moreover, the performance increases in the IBM when going from
- 100x100 to 1000x1000, while it typically decreases with the other
- processors.
-
- But on a classical molecular dynamics benchmark, the results are
- opposite! {DEC,HP,SGI} are almost twice faster than the IBM
- offering in the same price range.
-
- So we buy them all, for the desperation of our system manager,
- and run big electronic structure codes, doing mostly linear algebra
- or anyway very straightforward operations on huge loops,
- on IBMs with plenty of memory and no graphics display, and classical
- molecular dynamics, Monte Carlo and similar stuff on 32MB equipped
- (or sometimes even 16MB) workstations of other manufacturers,
- which are IMHO also nicer for interactive work.
-
- Apparently, one of the key issues is the cache architecture:
- size and bandwidth. The first code described above, doing linear
- algebra calculations, is very local in memory and uses a large
- and fast cache memory at its best. In contrast, the second code
- is highly non-local (it has lots of indirect addressing and jumps
- in memory like hell), and probably gives a poor Megaflops rating
- in all cases. Clearly, different design decisions have been made
- by the various teams. HP and SGI seem to behave quite similarly.
-
- Another issue giving a good Linpack rating for the IBM is the
- multiply-and-add instruction, where these two operations are
- done in a single machine cycle. This is clearly exploited, since
- the 84 MFlops result is obtained on a 50 MHz machine.
- IBM put some carefully coded BLAS routine in their ESSL
- library. Maybe that HP and SGI also have such an instruction,
- but it is not exploited by the compilers/libraries yet.
-
- DISCLAIMER: I am a poor application programmer trying to
- understand these wonderful machines "on the field", and make
- the best choices in the market. I really do not know what is behind the
- scenes ... Any insight and further input is appreciated!
- And, of course, number crunching is not all, as anybody struggling
- to land with the 747 on that damned short runway knows ... :-)
-
- --
- Furio Ercolessi
- Materials Research Laboratory | Intl School for Advanced Studies
- Univ. of Illinois at Urbana-Champaign | Trieste, Italy
- furio@uiuc.edu | furio@sissa.it
-