Subject: Long jobs - Steve Ellis report on SPARC 20 performance Date: Thu, 08 May 1997 17:48:22 EDT From: "Naomi B. Schmidt" ------- Forwarded Message Received: from SOUTH-STATION-ANNEX.MIT.EDU by po6.MIT.EDU (5.61/4.7) id AA05451; Wed, 20 Sep 95 16:31:15 EDT Received: from MITHRIL.MIT.EDU by MIT.EDU with SMTP id AA15481; Wed, 20 Sep 95 16:30:31 EDT Received: by mithril.MIT.EDU (5.0/4.7) id AA12631; Wed, 20 Sep 1995 16:30:50 -0400 Message-Id: <9509202030.AA12631@mithril.MIT.EDU> To: nschmidt@MIT.EDU Subject: Sparc Date: Wed, 20 Sep 1995 16:30:49 EDT From: Steve Ellis Content-Length: 1642 Naomi I have done a little more work on determining performance for a multi-processor Sparc. Here is a summary of what I've found: a) For cases where memory and swap are not factors, I didn't find any "slow down" due to numbers of jobs. b) For memory intensive jobs, performance begins to degrade when memory required is about twice the RAM installed. In the worst case, i.e. all memory and swap used, jobs got only 20 minutes of CPU time over a 5 day period. This was about 1% of the time available. For the "in-between" cases, performance becomes very erratic. Here are some "ps" numbers for jobs started together (18 total jobs): UID PID PPID C STIME TTY TIME COMD ellis 10542 10538 80 14:15:22 pts/0 37:53 ../linpackd ellis 10532 10517 80 14:15:18 pts/0 16:00 ../linpackd ellis 10530 10516 80 14:15:18 pts/0 8:30 ../linpackd ellis 10536 10522 80 14:15:19 pts/0 4:31 ../linpackd ellis 10533 10519 80 14:15:18 pts/0 2:29 ../linpackd ellis 10546 10537 80 14:15:22 pts/0 1:02 ../linpackd As you can see there is a wide range of values for CPU time. The machine does not appear to handle large scale swapping gracefully. For reference, the jobs I was running required about 16 Mb of memory, similar to what ADINA might take. The machine had 2 processors and 128 Mb memory. From this I'd say that even with large amounts of memory, a Sparc based batch server would be extremely susceptible to thrashing unless there is active operator intervention or management. Steve ------- End of Forwarded Message