home *** CD-ROM | disk | FTP | other *** search
- Xref: sparky comp.protocols.nfs:1924 comp.sys.sun.hardware:3479
- Path: sparky!uunet!sun-barr!news2me.ebay.sun.com!jethro.Corp.Sun.COM!exodus.Eng.Sun.COM!tabitha.Eng.Sun.COM!beepy
- From: beepy@tabitha.Eng.Sun.COM (Brian Pawlowski)
- Newsgroups: comp.protocols.nfs,comp.sys.sun.hardware
- Subject: Re: NFS I/O Ops/seconds
- Date: 22 Jul 1992 16:59:11 GMT
- Organization: Sun Microsystems Inc., Mountain View, CA
- Lines: 88
- Message-ID: <l6r4uvINNf0p@exodus.Eng.Sun.COM>
- References: <1992Jul22.061146.15641@u.washington.edu>
- NNTP-Posting-Host: tabitha
- Summary: nhfsstone PRE-LADDIS iops nfs
-
- In article <1992Jul22.061146.15641@u.washington.edu>, kint@rio.engr.washington.edu (Rick Kint) writes:
- > The Sun rep commented that a typical Ethernet allows 300 NFS ops per second,
- > so if you're on a single wire any server bandwidth beyond that is wasted.
- >
- > I have no feel whatsoever for the numbers that go into this figure,
- > can anyone in netland (a) explain this in words of two syllables or less, and
- > (b) does anyone know the figures for current systems?
-
- I decided to post this in addition to mailing. I have
- far exceeded two syllables. Sorry.
-
- The number "300 NFS ops/s per ether" is a rule-of-thumb
- (that's not quite the right word) maximum NFS ops/s as
- generated by nhfsstone (Legato) or PRE-LADDIS (SPEC)
- benchmark. I'll assume an equivalence of "NFS ops/s"
- and "NFS I/O ops/s". [Is there a difference?]
-
- It assumes a mix of which the more interesting features
- are:
-
- 34% Lookup
- 22% Read
- 15% Write
- 13% Getattr
-
- And something like 85% of the write's are 8KB, the amount
- of data touched is in excess of 15MB (dependent on number
- of clients). Couple hundred files, etc. All having implications
- on effectiveness of caching under the presented load.
-
- I'm truly generalizing above, by putting nhfsstone and
- PRE-LADDIS in same breath. From a 1,000 feet they're similar
- enough to generalize about.
-
- The "300 ops/s" number is approaching ethernet maximum
- utilization. For nhfsstone from Legato, I have faint
- recollections of this being in excess of 50-60% ethernet
- utilization with peaks going higher (as shown on a
- Sniffer in monitor mode).
-
- nhfsstone is the original benchmark from Legato. SPEC is
- evaluating a derivative benchmark called PRE-LADDIS. If
- you were going to get serious about benchmarking and comparisons,
- then you have to be precise about which benchmark
- you're talking about. On one recent version of PRE-LADDIS
- (each version has been changing during the "beata test",
- recent changes have been significant to load characteristics)
- I saw about 330 ops/s on an ether. Someone told me they
- saw 340 ops/s on an ether. At this point you have very
- busy ethernet.
-
- What do you do with this information? That's the hard part.
- How closely does your network resemble the PRE-LADDIS
- workload model? Probably not very. If you assume it
- does, and your load peaks are 300 and below, you can
- convince yourself that you can get by with a server
- with capacity to handle 300 ops/s and one ethernet
- interface (no need to segment net and have multiple
- interfaces to feed server). But doing an analysis of your
- network is not simple, and applying nhfsstone results
- to draw conclusions for your workload is not trivial
- (that is my personal opinion).
-
- Both nhfsstone, and more so PRE-LADDIS, are configurable and allow
- you to simulate other workloads. So for the serious
- investigator, you can conduct experiments to compare
- different server configurations for your specific workload.
- After you have characterized your workload to your satisfaction
- such that you feel your conclusions will be valid.
-
- I'm not sure I answered your "why 300 ops/s question" yet:-)
-
- It's just an observed point on or about which ethernet
- approaches maximum utilization when subjected to a load
- generated by PRE-LADDIS or nhfsstone (Legato).
-
- Both benchmarks are available (the PRE-LADDIS one is an
- evaluation version from SPEC with restrictions on use of
- data, etc). In the last InterOp proceedings there are
- slides describing LADDIS work which could explain some
- things.
-
- You asked for numbers for various servers... Can't do that,
- and the PRE-LADDIS beta evaluation agreement specifically
- precludes public disclosure of results.
-
- Brian Pawlowski
-
-