Benchmarking a 4x1.5Ghz Itanium 2

[Mephistopheles]

Here's something interesting...

...there's a lot to think about Itanium. It can destroy in SPEC; however, its acceptance has been lousy and adequate reports on its true performance are rare. What we know about it is, let's face it, dim at best.

Now here's the catch: the physics institute in which I work is about to receive a test system from HP using either a quad- or maybe a dual-Itanium 1.5Ghz, 6MB (Madison) configuration. This is only a test system to see if there is any interest at all in buying the thing.

In any case, what we need is an adequate way of benchmarking the thing. We have our heavy-on-FP programs (typical scientific stuff...), and we'll try those, but I was thinking about making a more thorough evaluation of Itanium's performance, since I'm about to have access to what is theoretically one of the best Itanium configs possible... Before the release of the newer models this year.

Any thoughts on what programs to recompile/run? There's always the open-source POV-Ray, and... let's see... (open source is better for benchmarking because we can recompile appropriately for Itanium)... If recompiling and learning how to use that is too hard or doesn't truly reward us with good performance, well, then...

Just as an experiment...

<i><font color=red>You never change the existing reality by fighting it. Instead, create a new model that makes the old one obsolete</font color=red> - Buckminster Fuller </i><P ID="edit"><FONT SIZE=-1><EM>Edited by Mephistopheles on 05/13/04 02:48 PM.</EM></FONT></P>

 

[juin]

Madison is usefull if the working set of the programe is very large.A benchmark that just use the same 512 KB is not very usefull for a very large cache cpu like madison.

It maybe be good also for your departement to think about the 1.6 ghz 3MB cache version is your set is smaller.

Also do you have access to HP compiler it do a very good job for HP unix.All benchmark that HP release allwayse use Hp unix with ICC or HP compiler.Often library are compile with ICC and HP compiler.Also flag and PGO offer large performance increase for I2 better to use those.

i need to change useur name.

 

[Kelledin]

Hmmm...you might try MATLAB (IIRC it runs on Linux).

Also, see if you can get the Intel C/C++ compiler for IA64. It should be free for non-commercial usage. Whatever you decide to recompile, try benchmarking them built-with-gcc against built-with-icc.

Sadly, icc isn't a perfect drop-in replacement for gcc. You may need to tweak some source code to get your favorite package to build with icc. I might be able to give you a hand with that, if you set up an SSH login.

<A HREF="http://lbs.sourceforge.net/" target="_new">Here's</A> a semi-official list of Linux-ready benchmarks. Just at first glance, BYTEmark might do what you want.

<i>Look! Up in the sky! It's a bird! It's a plane! It's...

...an asthmatic werehamster?

<LHGPooBaa> Well, @#!& on me.</i>

 

[Mephistopheles]

Thanks for the answer, juin. We are going to get a HP-UX OS with it, but I've heard a few complaints about HP-UX. It would seem that Linux would be a better choice, but we'll see about that. We don't have a choice, If I got that straight, anyway.

As for the 1.6Ghz, 3MB version, I think it might even be a smarter idea, but we don't have much of a choice about our test system. What I do know is that they will get us a juicy system (several gigabytes of memory, quad-processor configuration, heavy-duty HDDs, and so on), but we don't have that much of a choice.

What I plan on doing, anyway, is try to ascertain the potential this platform has. If it has a lot, we might end up changing the configuration at hand before we buy an I2-based system. We've got all tools available for benchmarking, including ICC, HP compilers, and so on. Many people around here also enjoy using FORTRAN as a scientific language; so it's also something we will be considering.

<i><font color=red>You never change the existing reality by fighting it. Instead, create a new model that makes the old one obsolete</font color=red> - Buckminster Fuller </i>

 

[Mephistopheles]

Thanks for the answer too, Kelledin.

Um, MATLAB is not an open-source program, and it would be best if we could recompile the programs we'll be using. But I'll look into that.

Comparing different compilers is also a nice idea. How well does HP-UX and I2 do at 32-bit code? What is it capable of running, and what is it not capable of?.... I wonder if I can answer some of those questions... BYTEmark is a nice one!

It would be great if I could find benchmarks for which I have numbers from other platforms... Maybe that would enable a more thorough Opteron/Itanium comparison, who knows...

<i><font color=red>You never change the existing reality by fighting it. Instead, create a new model that makes the old one obsolete</font color=red> - Buckminster Fuller </i>

 

[juin]

I hear very good thing about fortran on itanium that something that should be test.

to find a fair benchmark that can be use all around except SPEC there is not much.

i need to change useur name.

 

[juin]

Also ask intel about there excution layer software for hp unix.So you can run X86 benchmark on it and see what come out of it.As strange as is look.Some X86 compiled run faster on itanium that Ia-64 due to smaller footage.

i need to change useur name.

 

[Kelledin]

Comparing different compilers is also a nice idea. How well does HP-UX and I2 do at 32-bit code? What is it capable of running, and what is it not capable of?

I'm pretty sure there are some 32-bit versions of HP-UX, but they would be old versions for old non-64-bit PA-RISC arches. The IA64 HP-UX is likely going to be 64-bit only, what with IA64 being "pure" 64-bit.*

If you're willing to settle for "emulated" 32-bit, I suppose you could see if Intel's IA32EL will run on that thing. I'm not sure how much good it will be, though, unless it's a full Bochs-style machine emulator. You could always resort to Bochs and run any x86 O/S you want, except that Bochs is SLLLLOOOOOWWWWW.

Another thing I thought of...you might try LAME as a benchmark, or one of the DivX encoders for Linux. Either one can be pretty heavy on floating-point.

*(gcc supports -m32 to generate 32-bit code on some 64-bit arches--but not on IA64 or Alpha.)

<i>Look! Up in the sky! It's a bird! It's a plane! It's...

...an asthmatic werehamster?

<LHGPooBaa> Well, @#!& on me.</i>

 

[HardWareBoss]

Sweet quad Itanium 1.5Ghz, 6MB Madison. Have fun playing with that baby. If you guys don't like it. I'll be glad to take it off your hands for you. :lol: