Itanium Superdome

[generic_1]

I was wondering if anyone is running Superdomes with Itanium and others with PA-RISC and what performance differences you have seen if any when running your applications on them.

[Michael Tully · Accepted solution]

We are running a test between a rx7620 and an rp8400. If we get a 50% performance increase, we will be purchasing. The rx7620 we have for evaulation has 4CPU's and 10Gb RAM. The rp8400 has 6CPU's and 10Gb RAM.

[generic_1]

Did you have to recompile your code for it to work or to tweak it to get the performance increase. Is it a database like Oracle or some inhouse code perhaps? Just curious I have wanted the chance to play with Itanium :).

[Michael Tully]

Hi Jeff,

The rp8400 has peopleslop and oracle 9.2.0 currently. We are setting up the same on this itanium server. The oracle is certainly a different port, but the peopleslop will be the same except for any CObol related programs, which will be re-compiled.

From what I can tell is a lot faster so far, especially the time to reboot.

Regards
Michael

[Steven E. Protter]

If you are making a purchase decision, I have been told the following by several people inside HP.

1) PA-RISC is something HP plans to phase out. Apparently it costs too much to develop each new chip set and there are two few units to spread the cost out over.
2) Most of the new database performance tests are being run on Itanium boxes. Check the trades. There is probably a reason for that.
3) My suppliers are offering me discounts up to 40% on Itanium boxes when compared to similar PA-RISC boxes. We're endlessly getting quotes for an offsite DR server. HP is having trouble selling them? Perhaps.
4) HP plans a unified code release of HP-UX later this year. The same code is supposed to run on PA-RISC and Itanium boxes.

I have been careful here to quote HP and not speak for HP. I believe there are enough statements in the public domain to support this view.

So the road ahead is somewhat muddled. I'm eager for the results of Mr. Tully's tests and envious of the powerful hardware he gets to run.

SEP

[Alzhy]

We are provide IT services to the bread-butter business of on of the giant telcos. We are about to cut over an entirely new core application in about a month's time.

However, we are still sticking with PA-RISC cpu's and in fact bravely refreshing our PA8700+ systems (rp8400/SuperDome) with PA-8800 (dual-core) systems (rp8420's and upgraded SuperDome)to increase CPU density. We've been doing this for the last month PRIOR to production -- and I must say I am totally impressed with the transition from the PA-8700+ to the PA-8800 systems and the continued remarkable stability and resiliency of HP PARISC systems. (which will probably largely due to US admins religiously following install guidelines and patches..)

We should be an all-PA8800 shop by the time the last PA-RISC system is deployed (rp7420's will join the mix soon). These systems are purportedly Itanium ready as well should we really need its touted processing prowess -- but for now, we're quite happy with PARISC..

[Ted Buis]

HP also has a pa-8900 under development. The pa-8800 is not the last of the pa-risc. Since the pa-8800 will run 11.11, there are no issues with recompiling for native performance. The pa-8800 based systems use the Itanium 2 bus and the same chip sets as the Integrity (Itanium 2) servers. The 3X faster path off the chip, coupled with the 32MB L2 cache makes the pa-8800 much faster than just the clock rate increase. The dual core improves greatly the performance density in the rack. The pa-8800 is a big boost for customers with applications not ready for the HP-UX 11.23. All that said, for natively compiled application, the Integrity Servers have better overall performance and better price/performance, and that was only strenghten by the May 4th announcement of the mx2 dual Itanium 2 processor modules, that put two Itanium 2 chips and a 32MB cache in the same space as the Madison module. So Integrity Superdome can now go to 128 processors, just like the pa-8800 Superdome. The mx2 module does have a lower clock rate (to avoid meltdown), so for technical applications the single processor Itanium modules still offer peak performance. So not only do you now have a choice between pa-risc and Itanium 2, but the single Itanium 2 module versus the mx2 dual Itanium 2 module. Note the mx2 is made by HP (with Intel chips), and is available only from HP.

[Stefan Farrelly]

Weve tested out superdomes and rp7410's running HP-UX/PA-Risc running Oracle against Itanium running Oracle and the PA-Risc servers were much faster with oracle benchmarks. Number crunching - C programs, were much faster on the Itanium - due to faster cpu clockspeeds (1.5Ghz v 875Mhz).
I think it will take more time for Oracle to optimize for Itanium.

[Robert-Jan Goossens]

Hi,

Not shore if you are intrested but I add it anyway.

http://www.cbronline.com/currentnews/84088db57bc626a780256e35003855e6

Regards,
Robert-Jan

[Florian Heigl (new acc)]

We're currently testing a small itanium box, but I don't think it will be used for more than that - testing.
On the other hand a 128CPU-Superdome is on it's way. HP surely seems to try to get away from it's PA-RISC line, but somehow I think they should someday get the point:
- Noone buys itanium for larger systems.
- Performane/Clockrate of the Itanium is 'not that good' for Enterprise Applications
- If they don't get enough chips sold, they should simply stop selling Itaniums. :)

Customers don't want to trade their PA-8800's against 1.8GHz Itaniums. They want 1.8GHz PA-8800's. As simple as that.

If HP drops the PA-RISC line, why shouldn't the customers drop HP? The PowerV will be faster than an Itanium anyway.
[...because IBMs support is horrible. Yeah. I know :)]

[Ted Buis]

Contrary to Stefan's experience with Oracle, the TPC-C numbers for Integrity Superdome (SD)using Oracle for the DB are dramatically higher with those with PA-RISC. Some of that is clearly due to greater RAM support on the Integrity SD tested (1TB) versus the PA-8700 based SD. Maybe Stefan can expand with more specifics on the testing that was done. Anything Integrity server running a pa-risc binary is not likely to perform as well as pa-risc at this point. There is still some pa-risc code in 11.23.

[Hein van den Heuvel]

Stefan wrote:

"PA-Risc running Oracle against Itanium running Oracle and the PA-Risc servers were much faster with oracle benchmarks"

That is not a typical experience.

Oracle has the potential to be significantly faster on Itanium than on PA-Risc. Several benchmarks HP did (TPC, Peoplesoft, Siebel, Sap) have confiurmed this (sadly, not all results could be published). There must have been something odd, probably fixeable, in the test setup if you observer anything else.

[you were not running an PA-oracle image on IPF were you? (if that is even possible... I never bothered to try).]

Stefan (and other readers),
do NOT accept an inferior result form Oracle on Itanium. Review your testing and if need be escalate!

Call in HP support, call your partner/ISV contacts or perhaps call HP pre-sales support if evauluating a potential new systems.

Certainly, if you feel so inclined, seek help in understanding dissapointing results in this forum also. But I'm afraid that in general too much background information will be needed to create a comprehensive problem statement, and you deserve more timely and more commited help then a best efforts here can offer. At the very leat we'd need the exact systems/models used, memory sizes, cpu speeds, SGA sizing, disk setup overview, rought nature of the test (read or write, batch or oltp, a few streams vs many streams, complex queries or singleton selects? TX manager? networked?...)

hth,
Hein.