Re: Performance advice

"Art S. Kagel" wrote: > > On Tue, 20 Apr 2004 03:46:36 -0400, Andrew Hamm wrote: > > > Andrew Hamm wrote: > >> sumGirl wrote: > >> > >> NETTYPE soctcp,2,150,NET > >> NETTYPE ipcshm,2,150,CPU> > > > I missed one. You have 3 CPUVP's allocated, so change the ipcshm to: > > > > NETTYPE ipcshm,3,100,CPU> > > > because then the load applied to the CPU will be better balanced. Art S. > > Kagel has expounded on this subject quite admirably in the past. I'd > > actually be interested to know if this point has been corrected in the > > engine code these days. Any thoughts Art? Any recent experiments applied to > > see if it still matters? > > I stand by my testing on this, Andrew. Nothing has changed, and likely > nothing will, it's in the nature of the beast. Your suggested changes to > NETTYPE are right on.> > Art S. Kagel I have a question in this context: We have 32 CPUs an run 28 CPVPs. (Solaris 7 , IDS V9.40.UC3W1) All connections come in via a serverside application bundling db requests for approx 3k users, partly still running on the database server: (but we wil be able to put that app on a different hardware very soon) 70 connection via ipc (from the very same machine the database instance runs on), 70 via TLI (from an other piece of hardware - 1Gbit connected), both groups of connections with a real lot of traffic. We started with NETTYPE ipcshm,11,??,CPU ## ?? dunno anymore & no access now and lowered this in 3 steps to now NETTYPE ipcshm,6,??,CPU with good results (and at the same time lowered tlitcp,12,??,NET to 7 now) Any explanation WHY 6 ipc pollthreads are better than 11? At first I actually expected the contrary of what we see..... much like Art tested & described it Any comments? dic_k -- Richard Kofler SOLID STATE EDV Dienstleistungen GmbH Vienna/Austria/Europe

Richard Kofler wrote: > I have a question in this context: > > We have 32 CPUs an run 28 CPVPs. (Solaris 7 , IDS V9.40.UC3W1) > > All connections come in via a serverside application bundling db > requests > for approx 3k users, partly still running on the database server: > (but we wil be able to put that app on a different hardware very soon) > 70 connection via ipc (from the very same machine the database > instance > runs on), 70 via TLI (from an other piece of hardware - 1Gbit > connected), > both groups of connections with a real lot of traffic. So you are saying that the 3k users go via a small number of connections (either independent or part of the 70+70 connections) plus the 70 + 70 connections? Therefore your connections would be approx 150 to 200? If so, then that's not really enough connections to actually demand so much load sharing from the engine. If all 3000 users were connecting independently, then you'd see the phenomenum (help captain) we are talking about. Or, maybe you've got TOO MANY CPU's. (Is that possible???)