We've got a project coming up we might be using Lisp on. We can pick any of
the major compilers and cpu's. Does anyone have the Gabriel benchmark
results for ACL or Harlequin on a current generation Alpha or Athlon? (We've
already got them for Intel.)
Jonathan Coupe
"Jonathan Coupe" <········@meanwhile.freeserve.co.uk> writes:
>We've got a project coming up we might be using Lisp on. We can pick any of
>the major compilers and cpu's. Does anyone have the Gabriel benchmark
>results for ACL or Harlequin on a current generation Alpha or Athlon? (We've
>already got them for Intel.)
Don't have numbers for you, but be warned that the Garbiel benchmarks
don't reflect the performance of an optimized application (that means
declared data and controlled data layout).
You should consider Ultrasparc as well. Compilers are very tuned and
the machines quite fast anyway.
Martin
--
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
Martin Cracauer <········@bik-gmbh.de> http://www.bik-gmbh.de/~cracauer/
"Where do you want to do today?" Hard to tell running your calendar
program on a junk operating system, eh?
> You should consider Ultrasparc as well. Compilers are very tuned and
> the machines quite fast anyway.
>
> Martin Cracauer <········@bik-gmbh.de> http://www.bik-gmbh.de/~cracauer/
By coincidence, I just got a mail on UltraSparc today on the Beowolf list:
On Sun, 5 Dec 1999, Mark Hahn wrote:
> for instance, the E10K is hardly a UMA machine,
> and it's 500 ns latency to memory is nothing to brag about. HP will
> sell you competitive machines. but besides, if you rewrite for
clustering,
> why the HELL would you bother with pokey old UltraSparcs, especially
> in a gold-plated mainframe like the E10K? a warehouse full of whatever
> Intel or AMD is shipping today would be a lot cheaper and faster...
> The price per processor of this beast must be 4-10x the price per
> processor of equivalent Intel processors and 3-4x the price of alphas
> that would deliver twice or better the per-processor performance. But
> there may be benefits that I know not of...
Just shows how much compiler-cpu fit can alter the relative performance of a
cpu. But I suspect we'll go with an Inteloid or Alpha.
I take your point about the Gabriel bm's weaknesses - thanks.
Jonathan Coupe
* Jonathan Coupe wrote:
> By coincidence, I just got a mail on UltraSparc today on the Beowolf list:
Yes, but you don't *necessarily* want to trust anything beowulf people
say, as they tend to conveniently neglect stuff like interconnect
bandwidth & so on. Anyone can make a multiprocessor machine go really
well for problems that need no intercommunication...
--tim
Tim Bradshaw <···@tfeb.org> wrote in message
····················@lostwithiel.tfeb.org...
> * Jonathan Coupe wrote:
>
> > By coincidence, I just got a mail on UltraSparc today on the Beowolf
list:
>
> Yes, but you don't *necessarily* want to trust anything beowulf people
> say, as they tend to conveniently neglect stuff like interconnect
> bandwidth & so on. Anyone can make a multiprocessor machine go really
> well for problems that need no intercommunication...
>
> --tim
Actually interconnect bandwidth seems a dominant topic on the group these
todays - perhaps things have improved? Anyway, I'm concerned with single
macine performance. Good point though.
Jonathan
"Jonathan Coupe" <········@meanwhile.freeserve.co.uk> writes:
>By coincidence, I just got a mail on UltraSparc today on the Beowolf list:
>On Sun, 5 Dec 1999, Mark Hahn wrote:
>> for instance, the E10K is hardly a UMA machine,
>> and it's 500 ns latency to memory is nothing to brag about. HP will
>> sell you competitive machines. but besides, if you rewrite for
>clustering,
>> why the HELL would you bother with pokey old UltraSparcs, especially
>> in a gold-plated mainframe like the E10K? a warehouse full of whatever
>> Intel or AMD is shipping today would be a lot cheaper and faster...
>> The price per processor of this beast must be 4-10x the price per
>> processor of equivalent Intel processors and 3-4x the price of alphas
>> that would deliver twice or better the per-processor performance. But
>> there may be benefits that I know not of...
Sorry, this doesn't sound like you could derivate Lisp performance
numbers from it, I even got the impression that this is a simple rant
by an uninformed person and/or applies only to very specialized
applications.
Newest Sun material is at least up to Alpha in price/performance
(harder to say for Intel/AMD) and the compiler difference can easily
eat this up.
Not to speak of conceptual differences. If you use the Alpha with 64
bit pointers and you run a Lisp system with heavy use of Lists, the
pointers building up this list with take twice as much space as on a
32 bit platform. You may be able to buy enough RAM, but you usually
can't do anything about Cache and the memory bandwidth is halfed when
you calculate pointers/second, not MB/second.
I don't say that Alpha is bad or anything, in fact I like Alpha/OSF/1
more that SPARC/Solaris (and I'm pretty near to a point where I need
64 bit pointers), but if you need a fast Lisp machine, you have to
test your application on it yourself.
If you are going to buy one of the commercial Lisps, they may be able
to grant you access to the machines in question.
If can't go through this process, I would just take Allegro on x86 (I
would prefer FreeBSD over Linux), since that platform is most
"maneuverable", you can exchange and upgrad parts easier.
Martin
--
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
Martin Cracauer <········@bik-gmbh.de> http://www.bik-gmbh.de/~cracauer/
"Where do you want to do today?" Hard to tell running your calendar
program on a junk operating system, eh?
Martin Cracauer wrote:
> If can't go through this process, I would just take Allegro on x86 (I
> would prefer FreeBSD over Linux), since that platform is most
> "maneuverable", you can exchange and upgrad parts easier.
What makes you think that? Debian is quite flexible and has a huge package
list. Try it out... It's cool :)
--
-> -=-=-=-=-=-=-=-=-=- < Rahul -=- Jain > -=-=-=-=-=-=-=-=-=- <-
-> "I never could get the hang of Thursdays." -Douglas N. Adams <-
-> -=-=-=- URL: http://hoohoo.brrsd.k12.nj.us/~rjain/ -=-=-=- <-
-> -=-=-=-=-=- E-mail: ·················@usa.net -=-=-=-=-=- <-
Version 9.105.999.1111111111111.23.042
(c)1996-1998, All rights reserved.
Disclaimer available upon request.