home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.protocols.nfs
- Path: sparky!uunet!darwin.sura.net!mips!odin!sgihub!zola!twilight!speaker.wpd.sgi.com!coolidge
- From: coolidge@speaker.wpd.sgi.com (Don Coolidge)
- Subject: Re: NFS I/O Ops/seconds
- Message-ID: <nkkvj9s@twilight.wpd.sgi.com>
- Sender: news@twilight.wpd.sgi.com ( CNews Account at twilight.wpd.sgi.com )
- Organization: Silicon Graphics, Inc.
- References: <1992Jul22.061146.15641@u.washington.edu> <64081@hydra.gatech.EDU> <1992Jul22.185748.852@zia.aoc.nrao.edu>
- Distribution: usa
- Date: Wed, 22 Jul 92 22:29:57 GMT
- Lines: 78
-
- Seems like some summary is in order.
-
- 1) The standard NFS benchmark is nhfsstones. It will probably
- soon be replaced by LADDIS. Both are similar; in fact, the
- default mix of operations used by each is (at least was) the
- same. Reads/writes account for less than 35% of the mix.
-
- 2) Sun (among others) has from time to time reported very
- high IOPS numbers by changing the mix to have even fewer reads
- and writes. This is not dishonest...it just means that the
- reporting requirements for the test should make some mention
- of the mix.
-
- 3) Auspex has gotten well over 1000 IOPS with the standard
- nhfsstones operations mix, but that was indeed with
- 4 connected Ethernets, and using their own NFS load generator,
- not nhfsstones (they may well have also gotten wonderful numbers
- with nhfsstones, but I haven't seen any).
-
- 4) LADDIS reccommends adding another Ethernet for every 200 IOPS.
-
- 5) However, that's not really necessary...our own tests have shown
- that nhfsstones are at least as client-bound as they are server-bound
- and media-bound. We've seen well over 400 IOPS over a single Ethernet
- using nhfsstones with the default mix and with UDP checksumming
- enabled, between a Crimson server and a single Crimson client. We
- also see well over 500 IOPS with the same setup, but with two
- Crimson clients on the single Ethernet. Multiple Ethernets
- give an additional boost. So does use of FDDI instead of Ethernet
- (a big boost).
-
- 6) To be able to make reasonable NFS performance comparisons, I
- think that at least the following need to be specified in the
- performance report (there may well be more; this is just off the
- top of my head):
-
- - Network topology (number of nets; medium being used;
- number of clients per attached net)
-
- - Server: Number and clock speed of CPUs
- Medium Interface model and revision
- O/S version
- Number, kind, and size of disks attached
- Is there an NFS accelerator, either external
- or intrinsic?
- Any other specialized hardware?
- How much RAM?
-
- - Clients: Number and clock speed of CPUs
- Medium Interface Board model and revision
- O/S version
- Any acceleration onboard?
- Any specialized hardware?
- How much RAM?
-
- - General: UDP checksumming enabled?
- Mix of operations used
- Benchmark - nhfsstones, LADDIS, other?
-
- I've seen too many bogus comparisons made where The Competition's
- server was tested with underpowered, under-RAMed clients, or where
- different versions of the O/S were used on different clients, or
- where the mix was different, or the test, or...
-
- LADDIS is supposed to allow us to compare apples and apples. But it
- seems to have gotten a tad too political to do so, as many of the
- items I mentioned above were suggested and rejected as part of
- the output report. Too bad.
-
- Anyhow, if you want to be absolutely sure of your vendor
- performance comparisons, run the tests yourself. There's
- just too much information that gets left out of reports -
- for all practical purposes, any such marketing performance
- reports are worthless.
-
- Don Coolidge
- coolidge@speaker.wpd.sgi.com
-
-