RE: More CPUVPs than processors

I cannot comment on version 9, but on 7 I've done this for a long time. My current settings are 6 physical cpu's, but 12 virtual cpu's. If you watched my Unix cpu queue, it used to be quite idle, so I over-tuned my virtual cpu's, and the Unix queue went up, and my Informix queue came down. Was a little nervous in the beginning breaking away from the golden rule as taught in the manual and the performance tuning course, but it worked, and I'm still running with these settings happily. This will probably also differ from opsys to opsys. I am running on Solaris 9. Dirk -----Original Message----- From: owner-informix-list@iiug.org [mailto:owner-informix-list@iiug.org] On Behalf Of Mark Sent: 08 February 2005 04:09 AM To: ilist Subject: More CPUVPs than processors Has anyone found a benefit to running more CPUVPs than processors with 9.4? I have been running some load tests on a dual CPU Itanium/HPUX machine with 8gb ram connected to an EMC disk array. The instance is configured with 2 million buffers, 127 cleaners and LRUs. Rather than configure the instance specifically for bulk loading I have left the instance configured for OLTP activity - in real life many of the jobs that run during normal hours are similar in nature. Instance has 4 dbspaces, 4 spaces used for indexes, 3 temp spaces and a rootdbs. KIO is enabled. All spaces are on the same set of disks [no control over that]. Indexes are created after the data has been loaded. During the load I am finding that the cleaners do not manage to keep up. The LRU min/max are 0.5 and 0.7 respectively. The modified queues will reach as high as 15% dirty. If I have the system configured with 2 CPUVPS there will be idle cleaners. CPU usage rarely exceeds 25-30%. Checkpoints will take many 100s of seconds. I have found that adding CPUVPS results in all the cleaners being active all the time. The queues are much shorter, rarely exceeding 4% and checkpoints generally last no more than a minute. The extra kio threads use about 20-30% of the total CPU time used by all the KIO threads. The results are very repeatable. I have dynamically added/removed CPU VPS during large table loads and seen the effect on the queues and checkpoints. Mark sending to informix-list sending to informix-list

I also had fear about breaking from the golden rule, but got much better performance on 9.40 on AIX 5.2 with 4 physical processors using up to 12 CPUVPs. I think the test, test, test mantra is appropriate here, and it's easy to do, since the CPUVPs can be added dynamically.