Re: i know cary millsap is super smart but...
Date: Thu, 29 Aug 2013 11:33:18 -0400
Message-ID: <CAE-dsOKKv81fNMAV0b=88AKN=xSAZKD_FQ-TGL7FTGcsA+rydw_at_mail.gmail.com>
This is probably not directly related to this question, but its close. On my previous project we migrated from HP-UX where I believe we had 16 CPU servers in development/test. We supported multiple DBs on these servers. We went to IBM AIX, where we had 2, quad core CPUs (so 8 virtual CPUs), however, each DB was on a separate server. This worked without issue except when we had to do bulk changes that involved 'parallel sql'. I had a bulk update that used parallel 4 for both reads and the update. This seemed to cause a race condition on the CPU. When we looked at the events, 99.9% of the events where parallel waits. I take these to be messages piped to and from the query controller to tell the slave sessions what to do. When I checked the number of sessions created, it was a total of 8 plus 1 query controller. I know you can sometimes get 2x the number of sessions for each parallel level (this was debated on here back in 2007/2008).
Now I know that the other servers had a lot more CPUs, but they had many DBs running on them that were very active. So it is not like we were using alot of parallel and nothing else was running on these servers. Some of us started looking into the AIX docs and it appeared that the architecture of their CPUs/backend were considerably different than the HP-UX architecture. This was never going to be an issue in production. I cannot remember how many CPUs we had in production, but it was much larger and we did not have issues with using parallel. This actually made doing builds/maintenance more of a hassle because we had to lower down our parallel level. The initial hope had been that if we put each DB on a different server, we would be more efficient, because there are times when some dev/test environments were hit harder than others and this was leading to performance issues in other DBs on the same server.
I think there may be more to it than the number of CPUs/virtual CPUs. There may be architecture differences between systems. That being said, I am NOT a hardware expert. Is anyone on here familiar with IBM Server architecture? I do not know the exact server we bought. I don't work there anymore so I can't check.
-- http://www.freelists.org/webpage/oracle-lReceived on Thu Aug 29 2013 - 17:33:18 CEST