Date: Fri, 23 Feb 1996 11:00:36 -0800 (PST) From: "JULIAN Elischer" <julian@ref.tfs.com> To: kuku@gilberto.physik.rwth-aachen.de (Christoph P. Kukulies) Cc: freebsd-hackers@freefall.freebsd.org Subject: Re: FreeBSD supercomputing Message-ID: <199602231900.LAA05940@ref.tfs.com> In-Reply-To: <199602231044.LAA20191@gilberto.physik.rwth-aachen.de> from "Christoph P. Kukulies" at Feb 23, 96 11:44:18 am
next in thread | previous in thread | raw e-mail | index | archive | help
>
> I would like to gather some thoughts here in this list
> if such a solution would be feasible, also under
> the point of view that a bunch of users (ypcat passwd | wc -> 811)
> have to be supported continously (physicists compiling fortran
> programs, expecting correct results and having no time to
> discuss IEEE fp peculiarities or libm.a strangenesses, or
> perhaps new pentium bugs) OTOH, vendor fortran compilers aren't
> always as reliable as f2c, par example.
>
> There is already a strong L*n*x community I have to cope
> with and I would also like to gather benchmark data and
> other arguments pro FreeBSD - I'm under fire anyway for two
> years now why "*we* make FreeBSD while the other side of the
> aile is running L*n*x". I always argued with better networking
> performance, reliability and such. But fact is that some students
> installed L*n*x on their notebooks because they didn't find
> their PCMCIA ethernet adapter supported by FreeBSD while it was
> listed in the L*n*x 'hardware compatibility' list. These are
> the little needle stings I'm expiriencing every day.
l*n*x has PATHETIC networking
talk to russel..
here's an extract from www.geli.com
Single Node Low-End Workstation Price/Performance
Workstation SPECfp92 est. cost SPECfp92/$K
HP Model 715/100 100MHz PA-RISC 7100LC 137.0 $16.2 K 8.5
SPARCstation 20 Model 51 50MHz SuperSPARC 84.8 $9.3 K 9.1
SPARCstation 20 Model 612 60MHz SuperSPARC 127.1 $13.0 K 9.8
IBM POWERstation 40P 66 MHz PowerPC 77.2 $8.7 K 8.8
SGI Indigo2 Extreme 200MHz MIPS R4400SC 131.4 $15.7 K 8.4
Pentium 815\100 81.8 $5.1 K 16.0
These results show that workstations are priced relatively closely, in terms of performance, except for the Pentium
P54C, which has almost twice the price/performance on SPECfp92.
Cluster Price/Performance
Things get more interesting when high performance networking is configured into an eight node cluster. The following
is an analysis of the price/performance of an eight node cluster assuming FDDI and 100Base-TX ethernet. The analysis
is incomplete, because the cost for switches and cabling is not considered. It is, however, representative of actual costs.
The second column in the following table is the single node SPECfp92 data, multiplied by 8 nodes. The second column
provides an estimated cost for the standard configuration, with the addition of an FDDI adapter. The next column gives
the price/performance for the FDDI case. The final two columns repeat the exercise for 100Base-TX.
Eight Node Low-End Workstation Price/Performance with High Performance Networking
Workstation SPECfp92 cost SPECfp92/$K cost SPECfp92/$K
FDDI 100Base-TX
HP Model 715/100 1096.0 $182.1 K 6.0
SPARCstation 20 Model 51 678.4 $ 93.7 K 7.2 $79.3 K 8.6
SPARCstation 20 Model 612 1016.8 $123.1 K 8.3
IBM 40P 66 MHz PowerPC
SGI 200MHz MIPS R4400SC 1051.2 $149.8 K 7.0
Pentium 815\100 654.4 $ 52.0 K 12.6 $42.6 K 15.4
As before, the eight node P54C-100 system is almost 2x better in price/performance for equivalent networking
performance.
>
> Could someone name me some university sites which are
> preferably running FreeBSD or BSD type Os's and for what reasons?
>
> To name again the main issues:
>
> - Highspeed networking (100 Mbit, ATM, FDDI)
well that lets LINUX out of the picture..
>
> - SMP, clustering , distributed processing.
SMP is no a cost effective addition.. it complicates the algorythms
for scheduling work when you have to schedule work in two levels.
>
> - Numbercrunching (FP issues)
P6 is apparently pretty blinding
>
> - Disk speed
we have the best..
>
> Opinions and advice welcome.
>
>
> --Chris Christoph P. U. Kukulies kuku@gil.physik.rwth-aachen.de
>
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199602231900.LAA05940>