Questions for Oracle 9.2.0.6 on GS1280s

[John Dobbs]

Hello,

I am trying to gather some information to help our sysadmin configure the Tru64 5.1B-3 BL 26 kernel for Oracle 9.2.0.6.


We are going to move our two PeopleSoft apps and Oracle databases to 2 GS1280s clustered together. Each GS1280 has four 1.15ghz CPUs and 32GB memory. The SGA for database A is about 5GB and the SGA for database B is about 2.6GB.

Each GS1280 will normally run a single instance of Oracle 9.2.0.6. and PeopleSoft. But if a node goes down then both instances will be running on the same node. We will not link Oracle with numa_on. We are planning to set ssm_threshold = 0 and use granularity hints since it appears Oracle does not support big pages.

Sysadmin tells me we have 4 RADs. Is this optimal or should we reconfigure to 1 or 2 RADs?

To use 16GB for shared memory, following the System Configuration and Tuning manual, I came up with this:

rad_gh_regions = 4096

gh_chunks = 1024

Would this be correct? Oracle install guide uses a different formula for gh_chunks so I am confused.

Did I read that gh_chunks is obsolete on 5.1B? Do you use rad_gh_regions and gh_chunks together?

Any other guidelines/recommendations?

Your expertise is appreciated.

Thanks,
John

[Mark Poeschl_2 · Accepted solution]

With a GS1280 each CPU is its own RAD by default. I'm not even sure if it's possible to set things up any other way. So to set aside 16 GB for shared memory for your SGAs you would set

rad_gh_regions[0] = 4096
rad_gh_regions[1] = 4096
rad_gh_regions[2] = 4096
rad_gh_regions[3] = 4096

You can just ignore gh_chunks on your system. This parameter is not obsolete on 5.1B - it just doesn't apply to anything other than RAD 0 in your H/W configuration.

Setting ssm_threshold to zero is the right thing to do if you're using the rad_gh_regions parameters. I'm not so sure about the numa_on for Oracle. The GS1280 is certainly "less NUMA" than its predecessors the Wildfire class machines (GS80/GS160/GS320). Has either HP or Oracle specifically made this recommendation? If you have the luxury to try it some performance testing might be in order.

[John Dobbs]

Thanks Mark for your input. Regarding our decision not to make Oracle NUMA-aware, we did not want to introduce too many changes at this time. A post from Han Pilmeyer also leads us to think the effort might not be worth the benefit:
___________________________________
Running Oracle in NUMA is something that is very important on GS80/GS160/GS320 platform. However on the ES47/ES80/GS1280, because of the much better memory latency and bandwidth, it is not necessary.

Running Oracle in NUMA mode will make it a lot more difficult to administer for a very tiny benefit (if you see an improvement at all). I yet have to meet the first customer that runs Oracle in NUMA on this platform. It might be something to consider for the most demanding high end boxes (GS1280 with 32 cpus and more), but not for the ES47.
____________________________________

So we decided not to do numa_on but I am open to comments, suggestions, etc.

One other thing: I need some clarification of the System Configuration and Tuning manual. If you look at sections 4.4.8.3, 4.4.8.4 and 4.4.8.5 about modifying rdg attributes, there seems to be a couple of typos. I think 4.4.8.3 is correct but the other two sections both reference the same max_async_req paramter. Can anyone clear this up for me?

Thanks,
John

[John Dobbs]

I think we've got it figured out. Thanks.