home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.protocols.nfs
- Path: sparky!uunet!snorkelwacker.mit.edu!stanford.edu!ames!sgi!rhyolite!vjs
- From: vjs@rhyolite.wpd.sgi.com (Vernon Schryver)
- Subject: Re: Free NFS Response Time Measurement Software
- Message-ID: <op4fvdo@rhyolite.wpd.sgi.com>
- Summary: you never get more than what you pay for
- Organization: Silicon Graphics, Inc. Mountain View, CA
- References: <1992Aug10.045921.13353@aim.com> <nxfyh!-@rpi.edu>
- Date: Wed, 19 Aug 1992 14:41:01 GMT
- Lines: 79
-
- In article <nxfyh!-@rpi.edu>, fitzgb@mml0.meche.rpi.edu (Brian Fitzgerald) writes:
- > NFS RTM Distribution writes:
- > >
- > > *************************************
- > > FREE NFS SERVER RESPONSE TIME UTILITY
- > > *************************************
- >
- > ... cool program. For anyone interested, here's an example output of
- > the free demo they sent me. Good teaser for the sharpshooter.
- > ...
-
- > NFS RESPONSE TIME REPORT
- > Failure Rate NFS Response Time (ms) NFS Load
- > Server Over 20ms Avg Max (ops/sec)
- > -----------------------------------------------------------------------
- > hammer 0% 2.9 3.1 0.0
- > darwin 0% 1.0 1.0 0.0
- > usenet 20% 10.9 67.6 0.9
- > mml0 0% 2.5 2.7 0.1
- > cds1 9% 5.8 27.6 0.2
-
-
- I'd be very surprised if the AIM test is a fraction as interesting as
- LADDIS, the multi-vendor effort within SPEC to come up with an NFS
- benchmark.
-
- Deciding what to measure is always hard, and particularly difficult
- with NFS. Most people do not know and do not know they don't know
- which operations are common between an NFS client and its server. The
- response time of the server depends radically on which operations you
- assume are common, and about their nature. Trivially, what block size
- do you assume? What about client or server caching? Both of those
- have 10X effects on read and write speed. What about evil async
- writes, which can have >2X effects? I would not trust a benchmark that
- did not spend a lot of time convincing me it's assumptions about client
- working set are reasonable; you can expect "response times" to vary by
- 100X depending on the size of the client working set.
-
- The hard problem with the operation mix is that reads and writes are
- generally not as common as one would thing. The exact operation mix
- for LADDIS has been controversial. They have switched to a synthetic
- load generator from using real NFS clients. I guess it was too hard to
- define a standard client. (It's well known that the speed of the
- clients is at least as important as the speed of the server.)
-
- Another controversy within the group working on LADDIS has been what,
- if any, single "figure of merit" or number the LADDIS benchmark should
- produce. I don't recall seeing mention of "response time" for a long
- time.
-
- I'd be very surprised if the AIM test is very valuable. I could be
- wrong, of course, especially since I have never seen the AIM NFS test.
-
- I don't make this attack lightly, but the history of commerical
- benchmarks is not pretty. If you think the system vendors are scum,
- then you should check the benchmark slime. Would it be fair to say
- that SPEC was formed as response to Dhrystone and the so called
- commercial CPU benchmarks?
-
- If I wanted an NFS test, and if Silicon Graphics were not already
- involved, I'd wait a few more weeks and then buy the tape of LADDIS
- from SPEC. (I think the SPEC CPU benchmark tape costs $200).
-
- If I wanted a reasonable benchmark from people who understand NFS that
- is free or next to free, I'd contact Legato for a copy of nfshstone (it
- has an extra "h" in there somewhere). It was the starting point for
- LADDIS and is the benchmark that NFS server vendors are currently
- competing on. If you call a vendor and ask for numbers, you'll
- probably be able to get nhfsstone numbers.
-
- If I didn't want to spend any money, I would simply get a copy of the
- Sun NFS validation suite. It generates several numbers, and was the
- de facto benchmark among NFS developers for years. It's numbers are
- not very good, but at least you get the source and can see exactly what
- it thinks it is measuring. Of course, you also get the source for
- LADDIS and nfsshtone.
-
-
- Vernon Schryver, vjs@sgi.com
-