Oracle8i performance on SuperDome and XP512

[Sridhar Bhaskarla]

Woodish,

Disk I/O is of the bigger concern in oracled systems.

Did you consider doing a striping at the system level?.

I have seen a huge difference between default setup (using the LUNs sequentially) and the stripe setup. I have created a graph here that shows different throughputs for different stripe sizes.

1. extent level striping with Extent size=4MB
2. LV stripe of 64MB
3. LV stripe of 128MB
4. LV strip of 256MB
5. LV stripe of 512MB
6. LV stripe of 1024 MB

You will definitely notice the difference.

We tried cache LUN. It didn't buy us much. Instead we can try Solidstate for redo logs.

-Sri

[Andrew S Babb · Accepted solution]

Hi Wodisch,

2 CHiP pairs running High Performance mode will use up the ports on the XP, so the only suggestion is to make sure you stripe over all the pairs, and the 2 FC's on the HP.

On the XP the default stripe size for the Raid-5 is 48K, and with 5 Raid groups, you will probably want to use a stripe size of around 256K. We are still doing our testing, but the belief is that less than 256K is not advised, and once you go over 256K, there is not much difference. Our testing though is on a more populated XP.

I am guessing that you will have 4 LUN's per Raid-5 group, of about 12Gb's each. You will not really have enough spindles to put logs out on there own disks, so I would suggest the Oracle WP on SAME, for all LV's. Do a search for SAME (Stripe And Mirror Everything) on the http://otn.oracle.com/deploy/performance/listing.htm website.

We have not used Veritas on HP yet as we are still at 11.0 as opposed to 11i. No comment available on Vertias therefore.

Re Oracle level, and Volker's comments, you can improve instance recovery speeds via the log_checkpoint_interval & timeout parameters, even if you have larger log files. BTW - there is no indication of transactional throughput.

Also, you should be using multiple members on the redo logs. This will provide a small degree of comfort, but I would not run a production database without archivelog.

You say you are using 8i, is this exclusive or OPS mode? If OPS, and you can move to 8.1.7 then consider OPFS, available from the OTN download links. Provides failover capabilities for the database and application connections.

[Wodisch]

Hello,

finally I have to admit that I have not got the results, Andrew was changing faster than I was able to get copies of his results, so I actually do not know what was the "winning" setup. I still suspect the VxVM to have spoiled some of the tips...
Anyway: Thank you all for helping!!!
Kind Regards,
Wodisch