home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.sys.hp
- Path: sparky!uunet!mcsun!sunic!aun.uninett.no!Steinar.Haug
- From: Steinar.Haug@delab.sintef.no (Steinar Haug)
- Subject: Re: Cluster performance
- In-Reply-To: xm9@sdcc12.ucsd.edu's message of 12 Jan 93 14:10:45 GMT
- Message-ID: <STEINAR.HAUG.93Jan12170828@delab.sintef.no>
- Sender: news@aun.uninett.no
- Nntp-Posting-Host: tosca.er.sintef.no
- Organization: SINTEF DELAB, Trondheim, Norway.
- References: <1993Jan08.203304.13728@ttd.teradyne.com> <C0pDG7.Jp3@cup.hp.com>
- <43379@sdcc12.ucsd.edu>
- Date: Tue, 12 Jan 1993 16:08:28 GMT
- Lines: 35
-
- Tony Burzio writes:
- > NFS traffic between our HP730 and SunSSII workstation seems to be
- > transmitted at 2 Mbits/second. I remember reading a long time ago
- > that Sun had a trick to "take over" the ethernet in a non-friendly
- > way to speed up ethernet transfers at the expense of other clients.
- > This was why Sun NFS server performance was always better than HP's
- > for a long time. Perhaps this trick has migrated to the HP now as
- > well? Some vendors said that this was because the NFS service was
- > part of the kernel on the Sun, and that HP had only recently put the
- > NFS portion in the kernel to get the speed improvement. Since
- > Ethernet tops out at about 1.5 Mbits per second because of timing
- > constraints, I find this hard to believe...
-
- Hmmm... sounds like another variant of the old rumor that Suns send Ethernet
- packets with *less* than the minimum delay according to the Ethernet spec. This
- has been claimed by several people, and always been refuted. Van Jacobson, who
- probably knows more about Sun Ethernet performance than most people, says he
- has *never* seen this happen. Also note that with most Ethernet controllers,
- the packet delay is fixed by the controller and *cannot* be changed.
-
- On the other hand, it *is* true that Suns will send an 8 kByte NFS block of
- data as 6 Ethernet packets with minimum delay (9.6 us), and there are some
- other vendors' systems (DEC has been mentioned several times) that do not
- handle this very well.
-
- Note that with modern workstations it's easy to get more than 1 Megabyte per
- second between two workstations on an otherwise quiet Ethernet. I have seen
- (measured) this several times myself, using ttcp between Sparcstation IPCs.
- You should be able to get close to this with NFS reads, but don't expect to
- get close to it with writes (since NFS writes are synchronous on the server).
-
- Steinar Haug, system/networks administrator
- SINTEF DELAB, University of Trondheim, NORWAY
- Email: Steinar.Haug@delab.sintef.no,
- sthaug@idt.unit.no, steinar@tosca.er.sintef.no
-