If this isn't exactly what you wanted, please try our Search (there's a LOT of techy and non-techy stuff here about Linux, Unix, Mac OS X and just computers in general!):
From: Bela Lubkin <firstname.lastname@example.org> Subject: Re: Comments on RAID configuration wanted Date: Fri, 10 Jan 2003 21:45:34 GMT References: <email@example.com>
<firstname.lastname@example.org> Jeff Liebermann wrote: > Methinks you are optimizing > the wrong area. SCO ships with the disk buffers (NBUF) set at what I > would consider to be overly conservative. That was fine when RAM was > cheap, but makes no sense in these days of commodity ECC RAM. May I > suggest you read up on past comments regarding tweaking NBUF and NHBUF > in comp.unix.sco.misc. > <https://groups.google.com/groups?as_q=NBUF&as_ugroup=comp.unix.sco.misc> > Methinks you will find a rather spectacular improvement in disk > performance with an increase in these. On small boxes, my > rule-o-thumb is about 30-50% of RAM goes to disk buffering. There's > apparently a 450MByte maximum (NBUF=450000). Also, make sure you have > a really good UPS as leaving live data in the disk buffers is not a > great idea.
The other night I quadrupled the performance of a machine by _reducing_ NBUF from 400000 to 7000. Until this is fixed, be aware that the NFS client code on OSR5 gives you a huge performance _penalty_ for large buffer caches. As long as your machine isn't doing a lot of NFS client I/O (accessing files "over there" via NFS), it shouldn't be a problem. Technobabble: every time the NFS client code opens a remote inode, it tells the OSR5 buffer cache to purge old references to that inode. The particular routine it calls walks linearly through every buffer in the cache. With 400000 buffers, that takes a bit of time. This particular application was reading zillions of small files over NFS, compounding the problem greatly. NFS client machines which don't actually do a lot of NFS file accessing should be fine. Accessing large files via NFS should be fine. Systems whose whole lives are devoted to reading small files over NFS may benefit from sharply _reduced_ NBUF. Hopefully this will be fixed in a future NFS update (won't be fixed in 507 out the gate). >Bela<
Got something to add? Send me email.