Skip to contentSkip to conversation

benchmarking results

Linux-ia64 Adriano Galano 4 participants 8 messages Apr 30, 2003 – May 5, 2003

lore
Share & tools
lore
#0 lore raw

Hello:

Someone could be provide SMP benchmarking results for 2-way to 8-way Itanium 2 systems?

best regards,

-Adriano

#2 lore raw
In reply to Adriano Galano
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Adriano Galano

Adriano> Hello: Someone could be provide SMP benchmarking results Adriano> for 2-way to 8-way Itanium 2 systems?

A lot of performance information for 2-4 way is available at:

http://www.hp.com/products1/itanium/performance/index.html

Note especially the links near the bottom of the pages: that's where all the background & details are.

Also note that the results are for McKinley (as of July-September last year, I think), so they don't cover Madison (aka Itanium 2 6M) or larger (beyond 4-way) HP systems yets.

	--david
#3 lore raw
In reply to Adriano Galano

Hi David and all:

Quoted 9 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

On Wed, 30 Apr 2003 12:54:12 +0200, "Adriano Galano"

<[email protected]> said:

Adriano> Hello: Someone could be provide SMP benchmarking results Adriano> for 2-way to 8-way Itanium 2 systems?

A lot of performance information for 2-4 way is available at:

Quoted 9 lines

http://www.hp.com/products1/itanium/performance/index.html

Note especially the links near the bottom of the pages: that's where all the background & details are.

Also note that the results are for McKinley (as of July-September last year, I think), so they don't cover Madison (aka Itanium 2 6M) or larger (beyond 4-way) HP systems yets.

We need to find some reference of SMP scalability on Linux scheduler. For 1, 2, 4 and 8 processor. How "lineal" is this scalability? Does Linux perform well en 4 or 8 Itanium 2 architecture? Somebody have experience of Linux scalability for more than 4 Itanium 2 (or Madison) processors?

Regards,

-Adriano

Quoted 1 line
	--david
#4 lore raw
In reply to Adriano Galano
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Adriano Galano

Adriano> We need to find some reference of SMP scalability on Linux Adriano> scheduler. For 1, 2, 4 and 8 processor. How "lineal" is Adriano> this scalability? Does Linux perform well en 4 or 8 Itanium Adriano> 2 architecture?

There is no single answer to the scalability question. It totally depends on what application you're running. SGI has shown that certain apps scale very nicely all the way to 64 CPUs on their Altix box. Kernel-intensive apps won't scale that far, but even there there is some evendice of very nice scaling. AFAIK, there is only a 4-way TPC-C number published for Linux:

	rx5670 (4-way 1GHz Itanium 2):	80494 tpmC

but the URL here:

http://www.hp.com/products1/itanium/performance/commercial/tpcc.html

shows a TPC of about 40600 for a 2-way machine (I believe it was running Windows). Given that the 4-way TPC-C number for Windows is in the same 80k ballpark, it suggests that Linux scales almost linearly for TPC-C.

Adriano> Somebody have experience of Linux scalability for more than Adriano> 4 Itanium 2 (or Madison) processors?

NEC and SGI already are shipping large Itanium 2 machines. Not sure what benchmarks have been run on them with Linux, though.

As far as HP machines are concerned: it's probably best to get in touch with a sales rep. He or she should be able to get you more info on forthcoming machines and perhaps could help you get the application benchmarked.

	--david
#5 lore raw
In reply to Adriano Galano
Quoted 7 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Adriano Galano

Adriano> We need to find some reference of SMP scalability on Linux Adriano> scheduler. For 1, 2, 4 and 8 processor. How "lineal" is Adriano> this scalability? Does Linux perform well en 4 or 8 Itanium Adriano> 2 architecture?

Quoted 6 lines

David Mosberger wrote:
There is no single answer to the scalability question. It totally depends on what application you're running. SGI has shown that certain apps scale very nicely all the way to 64 CPUs on their Altix box. Kernel-intensive apps won't scale that far, but even there there is some evendice of very nice scaling. ...

Hi Adriano,

As David points out, there's a lot of factors that can affect scalability results for a particular workload. If you haven't already, you might want to check out the following from the February 2003 issue of Linux Journal
magazine: http://linuxjournal.com/article.php?sidd40, which provides more background on the SGI Altix system along with some scaling and benchmark results.

Hope that helps.

--steve

#6 lore raw
In reply to Adriano Galano

Hi Steve:

Quoted 20 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

On Wed, 30 Apr 2003 18:17:45 +0200, "Adriano Galano"

Quoted 2 lines · <[email protected]> said:

Adriano> We need to find some reference of SMP

scalability on Linux

Quoted 2 lines

Adriano> scheduler. For 1, 2, 4 and 8 processor. How "lineal" is Adriano> this scalability? Does Linux perform well en 4

or 8 Itanium

Quoted 7 lines

Adriano> 2 architecture?

David Mosberger wrote:
There is no single answer to the scalability question. It totally depends on what application you're running. SGI has shown that certain apps scale very nicely all the way to 64 CPUs on their Altix box. Kernel-intensive apps won't scale that far, but even

there there

Quoted 1 line

is some evendice of very nice scaling. ...

Hi Adriano,

Quoted 11 lines

As David points out, there's a lot of factors that can affect scalability
results for a particular workload. If you haven't already, you might want
to check out the following from the February 2003 issue of Linux Journal
magazine: http://linuxjournal.com/article.php?sidd40, which provides more
background on the SGI Altix system along with some scaling and benchmark results.

Thanks for the reference.

Maybe in all the cases I can see NUMA architectures and not find references related to what kind of scheduler was used older sched, O(1) or HT? The limits of the bricks for SMP are 4 proc, how could I assume linear scalability for 8 proc having memory latency accesing the others bricks through R-Brick maybe?

Regards,

-Adriano

Quoted 3 lines

Hope that helps.

--steve

#7 lore raw
In reply to Adriano Galano
Quoted 2 lines

Maybe in all the cases I can see NUMA architectures and not find references related to what kind of scheduler was used older sched, O(1) or HT?

Hi Adriano,

On the SGI Altix system we've been getting good results with the O(1) scheduler.

Quoted 3 lines

The limits of the bricks for SMP are 4 proc, how could I assume linear scalability for 8 proc having memory latency accesing the others bricks through R-Brick maybe?

There are many factors that can affect getting good "scalability" beyond 8p (e.g., hardware design, kernel scalability, user v. kernel workload, user app scalability, I/O subsystem, etc) -- having a good hardware design with a fast system interconnect is essential.

--steve