summary of answers concerning G90 on HP735

 Dear Netters,
 Thanks to all of you who replied.
 Here is the summary of the answers I have received to the question:
 >>3- What do you think about running G90 ( or G92) on HP735 ?
 >>How much memory do I need for an optimal run ? How fast would it
 >>be compared to the CRAY2 (or how slow) ? Our lab is in position
 >>to choose between 4 HP735 and a 2-processors Power Challenge from SGI.
 >>I would very much appreciate
 >>your opinions and a discussion "HP against SGI". The answers
 would be
 >>summarized, of course.
 from Murray Smigel,
 >First a disclaimer, I work for a computer company, Convex, which
 >markets HP-735 clusters, usually in conjunction with one of
 >our C-Series vector machines as the file and queue server.
 >
 >We have done quite a bit of comparison of HP-735 vs SGI R-4000
 >machines and the answer is that the HP is 2-3 times faster then
 >the R-4000 for Gaussian 92.  The "power challenge" SGI will
 >use a somewhat faster chip, I believe it will be a 75 Mhz vs
 >50 Mhz chip, but it will still be slower then the HP.  I think
 >4 HP's are a lot more machine then a 2 head SGI.
 >
 >You will need at least 64 MB of memory on the HP's.  What kind
 >of runs do you do?  Direct scf, mp2, mp4??  If you do
 >stuff with a lot of disk i/o, you might want to specify the
 >"fast and wide" SCSI interface as the older SCSI is not the
 >fastest disk in the world.
 >
 >I do not have direct comparison timings for the Cray 2, but for
 >most cpu bound things, the HP will be about 1/4-1/3 the speed
 >of a YMP.  If you want, I can run some data sets for you if you
 >have comparison timings on the Cray 2.  Just email them to
 >me with an estimate of the time, memory and disk requirements
 >
 >One other issue, Convex sells a scientific library product,
 >MLIB which is quite useful for linking Gaussian 92 on the
 >HP-735.  It really speeds up the matrix multiply dominated
 >steps.
 from Mike,
 (I've lost part of the message and the full name as well as
 the e-mail address. I apologize
 and beg to be contacted again; I have few more details to ask for)
 >We have been running G90 and now G92 on our HP 750 (slower than the
 >735) for about a year now, and it works quite well. You should probably
 >have 64 MB memory, but 32 MB would do if you don't do anything else
 >on the workstation (especially not use the display to login).
 >We also try to keep about 300 MB free disk space, though many of the
 >more popular G92 options can be done with little disk space, but
 >some can not. Our 750 is faster than the Cray X-MP we used to have,
 >but not be a lot (the 735 would be almost 2x our 750).
 >
 >At a first glance, you are getting ~twice as much total computing power from
 >4 HP 735s that you get from a 2-cpu SGI, but G92 is (partially)
 >parallelized for SGI so you might get better single-job turnaround
 >on the SGI by letting both cpus work on the same job. You want to have
 >local disk on the HPs to avoid doing I/O over the network,
 >though just loading binaries from a single /g92 directory might still be
 >acceptable (though you are dead if that single system goes down).
 >
 from Doug Fox,
 >  We are currently working on getting the HP/735 fully supported with
 >Gaussian 92.  Gaussian 90 is not available.
 >
 >  In general we recommend at least 32MB of memory and 1GB of scratch
 >disk for Unix workstations.  This provides sufficient balance and
 >resources that you can use all the capabiiilties of Gaussian 92.
 >