Oracle RAC on Itanium II Servers

[Douglas Buxton]

There is some doubt by our DBA's about performance of Oracle RAC on Itanium II servers. Question: Can Oracle RAC scale out (i.e. multiple 4 way servers) and perform comparably to large SMP servers with 6 or 8 CPU's? Is there any data on the performance characteristics in a scale out environment running Linux?

[Steven E. Protter]

HP has run tests on Itanium, running HP-UX and Linux that show its scalability. There are some press relseases on the subject ont he main http://www.hp.com website.

I've seen articles in oracle magazine as well. Those lead me to belive that Oracle RAC scales out reasonably well.

Oracle's got some fluff PR pieces as well on http://otn.oracle.com and http://www.oracle.com

You might want to set up a test in your local HP peformance center. They actually let you try things out to make sure they are acceptable prior to a hardware purchase.

SEP

[Fred Ruffet]

Hi !

Remember that RAC aims at stability, not at performance. One cluster running to bi-processors will not be compared to a single machine running 4 CPUs.

RAC will scale rather good if you consider that your server will (approximativly) always be up. You will really have to play to make your DB down...

Fred

[Douglas Buxton]

According to HP, RedHat and Oracle, RAC is touted as being highly scalable, implying that the performance charactistics are improved because of the clustering. That is the big question. Scaling out (like in a VMS cluster) is much more efficient from a performance standpoint than scaling up (i.e. large SMP systems). Is RAC's scaling comparable (from an Oracle performance standpoint) to typical "true" clustering?

[Steven E. Protter]

Oracle says yes.

I'm skeptical. A trip to the peformance center is what I'd do. Or have oracle give you some references you can call.

They should be motivated to do so, because they'll get license revenue from you implementing RAC.

Our plans are not use this technology. Our clustering will be ServiceGuard, and the aim is reliability not load balancing or scaling.

For what its worth.

SEP

[Fred Ruffet]

SEP is right, but I don't imagine that Oracle will give you references : Does anybody wanting to scale their server as managed to implement the two solutions in parallel, compare, through the bad one ? That would make a lot of money, no ?
Generally, in a firm, there are people who don't want us to buy big machines just to run test, and then to abandon them :)

Regards,

Fred

[Hein van den Heuvel]

Fred, within HP do that 'all the time'. Sometimes out of our own interest (think TPC-C, TPC-H, SAP SD 3-tier), sometimes driven by a sale opportinuty.

It seems that every time (a little bit of personal opionion injected here) the cluster solution for the same power, is cheaper and more available and every time the big-honking-iron solution is pushed as the desired solution (sometimes by hp, sometimes by the customer). Technically the cluster is clearly the way to go for many IT problems. Emotionally the big box still wins out. Sad but true.
Yes I have done some of those test (not yet on Linux though). We did test 4*4 and it performed better than 1*16 with Oracle being the main load.
If you carefully study big SMP tests you will see that often thy are run as multipel small boxes anyway! Either with vpars, or with mpsched, or psets, whatever. This is notably the case for Numa aware OS versions (like Tru64 was/is).

Still, you can compare it with traffic and having to move people or stuff around. Sometimes it is better to have lots of little cars sometimes only a big honking truck/bus will do the job (notbaly long haul, and just 1 driver/system manager huh?)

Hein.



Sarcastically yours,
Hein.

[Ted Buis]

The TPC-C record for all systems is currently held by an ORACLE RAC cluster on 16 HP rx5670 Itanium2 servers with 4 CPUs each.
See http://www.tpc.org/tpcc/results/tpcc_perf_results.asp?resulttype=all
That should satisfy anyone, that it can scale. You could debate how much effort it took to get that performance, but the result is very impressive, beating a 64-CPU Integrity Superdome in performance and price performance. Remember that there are 16 systems to administer, but if you lose one it is only 1/16th of your capacity.

[Ted Buis]

Actually this was Oracle 10g, so if you have to use 9iRAC, that might be a different story.