Date: Mon, 21 Jun 1999 14:05:59 -0700 From: Mika Nystrom <mika@cs.caltech.edu> To: Richard Cownie <tich@ma.ikos.com> Cc: freebsd-smp@freebsd.org Subject: Re: SMP, 4GB RAM, 4x CPU Message-ID: <199906212105.OAA29913@varese.cs.caltech.edu> In-Reply-To: Your message of "Mon, 21 Jun 1999 12:37:47 EDT." <99062113061000.18239@par28.ma.ikos.com>
next in thread | previous in thread | raw e-mail | index | archive | help
Richard Cownie writes:
[..]
>I've never used an Alpha myself. From the benchmarks I've seen I believe
>they are substantially faster for floating point, and they have more memory
>bandwidth and bigger caches. So if absolute performance on floating-point
>intensive stuff is what you need, Alpha is worth considering. However,
>if price-performance is what you need, I believe a dual-Celeron at 500MHz+
>with 1GB dram for $3K is about the best you can do. As far as I know
>the dual-Alpha systems are still up in the $10K range. This may change
>next year. A couple of months from now K7 systems may also be interesting
>(though I think we'll have to wait a while for cheap SMP-K7 systems).
>
>Also there are rumours that Alpha performance is critically dependent
>on the compiler used - i.e. the DEC compiler on Tru64 Unix (is that this
>week's name ?) might give you 30% more performance than gcc. This is
>third-hand information, so may be false (or it may be true for Alpha 21164
>but not 21264, which I think does more dynamic scheduling ?). Be wary
>of the SPECint/fp results - these may be heavily influenced by the cache size,
>and maybe also wacky compiler options which are rarely usable in real life.
>
[..]
Hello everyone,
From personal experience---we use FreeBSD/SMP, mostly on dual PPro
systems with up to 1GB, as well as Mach.. err OSF.. err Digital..
err.. Compaq Tru64 on a couple of Alphas (21164 @266 MHz & @500MHz).
The new DEC cc (that is, /usr/bin/cc -migrate -tune ev5, etc etc
etc---DEC cc is actually two compilers in one) can give a very
large performance boost on the 164/ev5 as opposed to gcc; in fact,
on Fortran code (which requires either f2c or even more money spent
on DEC f77), I wouldn't be surprised if the performance benefit
could be 100+%. But your guess as to the 264 is probably correct.
The 164 is really an in-order processor, and takes a very large
(up to 300% longer runtimes) performance hit for improperly scheduled
code, and it also has a very long-latency memory system that places
a large premium on controlling your caching behavior.
We did some benchmarking a few months ago, and I believe that the
266 MHz alpha running an in-house version of SPICE was close to
the same speed as a 450 MHz Pentium III (this was at the time close
to the latest-and-greatest Intel chip racing a four-year-old Alpha),
and the 500 MHz machine (two years old) just blew away the P-III.
(Of course, DEC cc -migrate vs gcc; I believe the P-III was running
Red Hat, but that shouldn't matter much.)
All that being said, if you want large memory, I think a 264-based
system could be better price/performance than a Xeon-based system
(have you seen what the large-cache Xeons cost? sheesh!) But if
you're on a tight budget, Celeron, obviously :) (And K7 soon, I hope.)
Regards,
Mika Nystrom <mika@cs.caltech.edu>
Asynchronous Systems Architecture Project
Department of Computer Science
California Institute of Technology
To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-smp" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199906212105.OAA29913>