johan@dutnak2.tudelft.nl (Johan Haas) (06/03/91)
We are considering to replace our current Convex C1 with a mix of RS/6000 systems instead of upgrading to a Convex C2. I would like to hear if e.g. the model 550 is any good to serve as a machine for, say, 8 PhD. students (like myself). We do a lot of program development, but also do finite difference simulations that run easily 10 hours on our C1. I know that the MIPS/MFLOPS performance of the 550 is superior to that of the Convex, but there are rumours that the multiuser performance of RISC machines is very bad. Is this true? Another issue is the single precision Floating Point performance that --according to a lettr to Unix World-- can be worse than the double precision performance. We do most of our work in single precision. I would appreciate any comments Johan de Haas tel. +31 15 785188 E-mail: johan@dutnak2.tudelft.nl Lab. of Seismics & Acoustics Delft University of Technology P.O Box 5046, 2600 GA Delft, The Netherlands -- Johan de Haas tel. +31 15 785188 E-mail: johan@dutnak2.tudelft.nl Lab. of Seismics & Acoustics Delft University of Technology P.O Box 5046, 2600 GA Delft, The Netherlands
paul@uxc.cso.uiuc.edu (Paul Pomes - UofIllinois CSO) (06/04/91)
johan@dutnak2.tudelft.nl (Johan Haas) writes: >We are considering to replace our current Convex C1 with a mix >of RS/6000 systems instead of upgrading to a Convex C2. I would closely examine the overhead of administering a RS/6000 system. Why IBM gave us AIX instead of UNIX is beyond me. /pbp -- Paul Pomes, Computing Services Office University of Illinois - Urbana Email to Paul-Pomes@uiuc.edu
mccalpin@perelandra.cms.udel.edu (John D. McCalpin) (06/04/91)
>>>>> On 3 Jun 91 14:31:34 GMT, johan@dutnak2.tudelft.nl (Johan Haas) said:
Johan> We are considering to replace our current Convex C1 with a mix
Johan> of RS/6000 systems instead of upgrading to a Convex C2.
This is a good idea.
Johan> I would like to hear if e.g. the model 550 is any good to serve
Johan> as a machine for, say, 8 PhD. students (like myself). We do a
Johan> lot of program development, but also do finite difference
Johan> simulations that run easily 10 hours on our C1.
You should not have any trouble with such a small number of users if
you have adequate memory and swap space.
Johan> I know that the MIPS/MFLOPS performance of the 550 is superior
Johan> to that of the Convex, but there are rumours that the multiuser
Johan> performance of RISC machines is very bad. Is this true?
There is no reason to call the performance "very bad". The reviews
that I can recall were disappointed with the multi-user performance
because it was no better than other machines with (for example) half
the MIPS rating of the IBM. Unless your 8 users each insist on
running *lots* of active processes all the time and all at the same
time, you should not have trouble.
Johan> Another issue is the single precision Floating Point
Johan> performance that --according to a lettr to Unix World-- can be
Johan> worse than the double precision performance. We do most of our
Johan> work in single precision.
It is regularly observed that single-precision (32-bit) arithmetic
runs between 10% slower and 10% faster than double precision (64-bit).
The moral is that (unless you are out of memory) you can switch to
64-bit for "free", with resulting increases in accuracy and
reliability. But even if you run in 32-bit precision, you should not
expect to see a penalty of more than 10-15%, and you might see a
slight speed improvement --- it depends on how well the compiler can
overlap the precision conversions for your application with other work
that has to be done anyway.... Try one of your own applications and
see what happens....
--
John D. McCalpin mccalpin@perelandra.cms.udel.edu
Assistant Professor mccalpin@brahms.udel.edu
College of Marine Studies, U. Del. J.MCCALPIN/OMNET
rmilner@zia.aoc.nrao.edu (Ruth Milner) (06/04/91)
In article <1991Jun3.173646.25682@ux1.cso.uiuc.edu> Paul-Pomes@uiuc.edu writes: >johan@dutnak2.tudelft.nl (Johan Haas) writes: > >>We are considering to replace our current Convex C1 with a mix >>of RS/6000 systems instead of upgrading to a Convex C2. > >I would closely examine the overhead of administering a RS/6000 system. >Why IBM gave us AIX instead of UNIX is beyond me. Especially if you have to run, say, half a dozen computers instead of one. Don't underestimate the increase in the system manager's workload. -- Ruth Milner Systems Manager NRAO/VLA Socorro NM Computing Division Head rmilner@zia.aoc.nrao.edu
heiko@methan.chemie.fu-berlin.de (Heiko Schlichting) (06/05/91)
mccalpin@perelandra.cms.udel.edu (John D. McCalpin) writes: >>johan@dutnak2.tudelft.nl (Johan Haas) said: >> We are considering to replace our current Convex C1 with a mix >> of RS/6000 systems instead of upgrading to a Convex C2. > >This is a good idea. I'm not sure about this. I still miss a good batch queueing system on RS/6000 - one of the famous things on a Convex. And in case of low swap space AIX kills the process which had allocated most memory. So everybody can kill large batch processes from other users which are running since a long time. Not a very good 'feature' of AIX. Bye, Heiko. -- |~| Heiko Schlichting | Freie Universitaet Berlin / \ heiko@fub.uucp | Institut fuer Organische Chemie /FUB\ heiko@methan.chemie.fu-berlin.de | Takustrasse 3 `---' phone +49 30 838-2677; fax ...-5163 | D-1000 Berlin 33 Germany
rtp1@quads.uchicago.edu (raymond thomas pierrehumbert) (06/05/91)
> (questions about 550 vs. Convex, esp. multi-user performance)
I think that replacing a Convex with a cluster of 550's could be
a big win. I benchmarked the R6000 series against a number of
other number crunchers, and found that the performance is
pretty much as claimed. Code that is very vectorizable
but with no possibility of re-use of data may still do
better on a vector machine like the Convex though; code in
this category may be rarer than you think.
Multi-user performance? I ran four copies of a 2D fluid code, and
then did a big makefile. It worked fine. I think 8 users wouldn't
be any problem, as long as you have enough memory. If it gets
to be a problem, buy a couple extra 550's! For the price differential,
you can afford this vis a vis 1 Convex.
Gee, I wish IBM would come out with a four-processor shared memory
version of the R6000. Now THAT would be a really fine cruncher.
.
1