Re: Processor Affinity

In article <5daceu$sun@cssun.mathcs.emory.edu>, Jon C Vemo <jvemo@cyberspace.com> writes >At 08:52 AM 2/5/97 -0500, Steve Romankiw wrote: >>Our Sun 4000 server supports processor affinity. Because Sun supports >>this option, I have configured our On-Line instance to recognize this >>option. Our present user count is low, roughly 50 concurrent users. >>Beginning next week, we will be jumping to over 300 concurrent users. >> >>In your environments, have you seen and +'s or -'s related to using >>processor affinity? >> It can help since it reduces the flushing required in the memory cache of the "pinned" CPU. This should help to increase effective CPU performance due to elimination of reduce cache flushes. >>Below are setting taken from my onconfig: >># >># Our Sun server has four processors Sounds like a good contener for processor affintity. >># >>MULTIPROCESSOR 1 # 0 for single-processor, 1 for>>multi-processor >>NUMCPUVPS 3 # Number of user (cpu) vps >>SINGLE_CPU_VP 0 # If non-zero, limit number of cpu vps>>to one >> >>NOAGE 0 # Process aging >>AFF_SPROC 1 # Affinity start processor >>AFF_NPROCS 3 # Affinity number of processors> Do many other processes run on this machine? Remember the above setts mean 3 CPU VP's each one can ONLY run on it's assigned processor. i.e. CPU VP 1 CAN ONLY RUN on Phyical CPU one. CPU VP 2 CAN ONLY RUN on Phyical CPU one. CPU VP 3 CAN ONLY RUN on Phyical CPU one. This means that IF ANY OTHER PROCESS RUNS OF ONE OF THE FIRST 3 PHYISCAL CPUs THAT THE ASSOCIATED CPU VP WILL NOT BE ABLE TO RUN EVEN IF PHYSICAL CPU 4 IS IDLE. Reread that statment for a while and consider that any other process has a 75% chance of running on one of the first 3 physical CPU's and so stops the assocaited CPU VP from running. I would probably try AFF_NPROCS as 2 or even 1. Also in some multiple Physical CPU machines the kernel only runs on CPU 1 so I would generally try to avoid that CPU unless you find out otherwise. (I know I'm being over cautious but that's my style!). >Processor affinity has been available for Sequent, since the early/initial >DSA (7.x) I heard a rumour that Sequent helped to develop DSA so I am not surprised. Strange how the DSA team were situated in a new building right next door to the main building holding Sequent technical staff!! >ports. I have had some very strange behaviour issues using in some of the >early >versions. However some recent testing of an OLTP application under more recent Same hear - once had an early DSA 7 on a 2 CPU Pyramid. We set PDQPRIORITY to 1 and ran a query. The machine froze solid and even the console stopped responding to keypresses so we couldn't even break out into the PROM. The client called out a Pyramid hardware engineer and he got into the PROM. Apparently an oninit process was in a very tight loop trying to get a semaphore which was not available!!! >versions (7.20), have showed some promising performance improvements. > >Proceed with caution though, you can easily shoot yourself in the foot. >Review carefully >what your server is running/doing, and CPU utilitization and load. > >Jon >---------------------------------------------------------------------------- >--- >Jon C. Vemo "Life is like a dogsled team, if you ain't the >jvemo@cyberspace.com lead dog the scenery never changes." >---------------------------------------------------------------------------- >--- > -- David Williams