home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!europa.asd.contel.com!darwin.sura.net!wupost!usc!sdd.hp.com!uakari.primate.wisc.edu!relay!afterlife!dpkemp
- From: dpkemp@afterlife.ncsc.mil (David P. Kemp)
- Newsgroups: comp.sys.sun.admin
- Subject: Re: Kernel config - pseudo-device win128 (and other stuff)
- Message-ID: <1992Jul21.130934.16098@afterlife.ncsc.mil>
- Date: 21 Jul 92 13:09:34 GMT
- References: <1992Jul15.232035.29008@kodak.kodak.com>
- Organization: The Great Beyond
- Lines: 52
-
-
- In article <1992Jul15.232035.29008@kodak.kodak.com> dennett@sunshine.Kodak.COM (Charlie Dennett) writes:
- >Our 4/490 server is a real workhorse. It is the lab's main server,
- >has 18 disk spindles, server home directories and some other
- >filesystems for well over 100 users, and supports a half dozen
- >(and growing) X terminals. Many of the users have Sparcstations
- >but this 4/490 is used for long running CPU and I/O intensive jobs.
- >We've investigating how to squeeze more performance out of it.
- >We noticed that the kernel config file has an entry:
- >
- > pseudo-device win128
- >
- >We were wondering if we had anything to gain by increasing the 128
- >to, say, 256? Any ideas? We already have maxusers set to 128.
- >
-
- Not unless you have a Sunview user sitting at the 4/490 console with
- *lots* of windows open :-). Since you are using X terminals, that's
- not very likely.
-
- >
- >We've also used adb to increase the buffer cache to a max of 112
- >as recommended in the O'Reilly "System Performance Tuning" book and
- >some other papers we've read. (Why is 112 the max?)
- >
- >Any other recommendations on squeezing more oomph out of this beast?
-
- Tuning for performance doesn't necessarily mean "more of everything",
- although since you have 128 M, increasing the kernel size with
- MAXUSERS can't hurt.
-
- I would focus in a different direction though - try to offload the
- server by running the long CPU-bound jobs on your Sparcstations
- (you didn't say SS-1, 1+, or 2, but even the 1 is a reasonable
- compute engine compared to the 490). Try to identify which of
- your jobs do lots of continuous disk I/O, and keep them on the
- server. Jobs that do sequential reads from a large file, and
- not much writing are ideal candidates for migration to the Sparc-
- stations.
-
- NFS and heavy computing just don't mix very well on an NFS server -
- the nfsd's get starved no matter how many of them you use, and the
- clients get "server not responding" and "server OK" message pairs
- out the wazoo.
-
- Your Sparcstations are "dataless", aren't they? If they are diskless
- (root and swap on the server) you have to go out and buy them local
- disks *right now*, no matter how tight money is.
-
- --
- Dave Kemp dpkemp@afterlife.ncsc.mil "Pave the Bay!"
- #include <std/disclaimer>
-