SOFPJF@VM.UOGUELPH.CA (Peter Jaspers-Fayer) (10/06/90)
Background: We are running an SGI 380, IRIX 3.3.1. The system supports a number of scientific researchers who (as yet) do no graphics, but crunch a LOT of numbers. The 8 CPU's are pegged at 100% almost all the time. I note that "sysview"s report of "cntxt" switches is very high at times. (8192, sample rate of 30 sec) with only 12 or so real CPU hogs running. We hope to offload some of this onto a 340. The questions: 1) How CPU intensive is context switching in IRIX? SAR reports very little (2-11) %sys, so my guess is it's not bad so far, but I sus- pect we will be getting more contention for the CPUs later. 2) What are the effects of taking the top 2 or 3 (according to TOP) processes and doing a `npri -t 300` on them? I note a drastic drop in context switches, but how does it (or will it) effect overall thru-put? Other than (ugh) benchmarking, I wonder how one could tell? 3) Where is the default time-slice documented? Does Unix in general or Irix in specific adjust the size of this on-the-fly, giving longer- running jobs longer cracks at the CPU? (I know of other OS's that do.) 4) We are evaluating/installing NQS (SGI batch scheduler). Does anyone out there have any idea: What is the maximum number of CPU-intensive processes one should run on an 'n' processor CPU? (n=8 or 4 in our case). (I know: "That depends", right? On what please?) 5) When (if ever) would one want to get into locking a process onto a specific CPU of a multi-processor system. /PJ SofPJF@VM.UoGuelph.Ca (Probably also reachable (until ?) at SOFPJF@UOGUELPH.BITNET) It is impossible to make anything foolproof because fools are so ingenious.