Fellow Sun-Managers,
I thought I'd wait a couple of days to summarize what I have been
sent, and since I haven't received any mail on this subject today, I
thought I'd summarize.
There were two basic comments:
1. Get nfswatch 3.0
2. Get nhfsstone
Now, nfswatch 3.0 is freely distributable, but I don't know that much
about it. Presumably it gives you more of the detailed ``10% of your
total procedure calls were write() calls, but they took up 40% of the
total time used by NFS'' type stuff I was looking for.
Legato Systems (maker of the PrestoServe VME and S-Bus NFS write cache
accelerator boards) supplies (apparently free of charge) nhfsstone, which
even Sun MicroSystems has taken to using in their own ads (according to
Jeorg Houck [*]). However, as I understand it, nhfsstone is more of a
benchmarking type of program (set it up and it loads your servers 'till
they are swamped, and then it tells you how far it got), as opposed to the
non-invasive historical observation-type of program that I was looking for
(similar to nfsstat). Interestingly, according to the Archie server at
quiche.cs.McGill.CA, both nhfsstone and nfswatch 3.0 are available via
anonymous ftp from numerous sites. I will leave it up to the reader to
locate the site nearest them that has the appropriate program they want.
One interesting comment from Hal Stern, Sun MicroSystems East, is that
nfsstat itself can supply some information of this kind, but only for the
simpler types of system calls (getattr, read, write, etc...), and only
under SunOS 4.1 and later. To quote from his reply:
>this is easy -- on your nfs clients (running 4.1 or later)
>run nfsstat -m. this gives you client-perceived server response
>time for "simple" requests - getattr, lookup, readdir - and
>for reads and writes. your simple requests should be *very* fast,
>but you may be going to disk to satisfy them if your server isn't
>configured ideally. for example, if your inode cache is really
>small, then you'll spend a lot of time pulling inodes off of disk
>to satisfy getattrs. similarly, directory reading will kill you
>if the server's DNLC isn't big enough. bump up MAXUSERS fairly
>high (64 on a 32M machine; 128 on a 64M machine).
Finally, many people told me that the general rule of thumb on buying
something like the PrestoServe NFS accelrator products is that if your NFS
writes are taking more than 5-8% of the total number of calls, then you
should probably consider one for your Server. If you care to look on the
back of the June 1991 issue of _The_Sun_Observer_, you will note an
advertisement from Interphase Corporation quoting a study done by Sun
MicroSystems, that shows that with their NCP board (handles Ethernet,
TCP-IP and NFS on a co-processor board, as opposed to caching just the
writes from NFS like the Legato product), you can get even better
performance. Well, I've seen parts of the study, and that is only true if
you have your server sitting on more than one Ethernet segment. In fact,
one or two of the Interphase NCP boards plus the Legato board seems to get
the best performance possible if your server is only on one or two
segments.
If you are interested in talking about this subject some more, either
send me private e-mail, or let's take this conversation over to the
Sun-Nets Mailing List, as I think the moderator for Sun-Managers, William
LeFebvre (who just happens to occasionally write for _The_Sun_Observer_,
and does so quite well, I might add) probably won't think too kindly of us
if we start caryying on a conversation here.
Thanks to the following (in random order):
lmcstst@noah.ericsson.com (Stamos Stamos)
sbradley@iwarp.intel.com (Seth Bradley)
stern@sunne.East.Sun.COM (Hal Stern - Consultant)
datri@concave.convex.com (Anthony A. Datri)
pochmara@cse.ogi.edu (John Pochmara)
jan@eik.ii.uib.no (Jan Henriksen)
pren@BBN.COM
librainc!ho@uunet.UU.NET (Alan K. Ho)
davee@tafs.mitre.org (Edwards, D N)
mitch@hq.af.mil (Mitch Wright)
synergy!ronin!kevin@Sun.COM (Kevin Sheehan)
chs!jeorg@boulder.Colorado.EDU (Jeorg Houck)
hanson@calvin.fnal.gov (Steve Hanson)
________________________________________________________________________
| Brad Knowles | Internet: blknowle@frodo.jdssc.dca.mil |
| DISA/DSSO/JNSL | Ph: (703) 693-5849 Fax: (703) 693-7329 |
| The Pentagon, Room BE685 |-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-|
| Washington, D.C. 20301-7010 | Speaking from, *not* for DISA (nee DCA) |
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
[*] Also according to Jeorg, you should be able to get it if you send
an e-mail message to nhfsstone-request@legato.com and ask them for
it.
This archive was generated by hypermail 2.1.2 : Fri Sep 28 2001 - 23:06:15 CDT