home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.protocols.nfs
- Path: sparky!uunet!gatech!darwin.sura.net!haven.umd.edu!decuac!hussar.dco.dec.com!mjr
- From: mjr@hussar.dco.dec.com (Marcus J. Ranum)
- Subject: Re: NFS corruption discovery
- Message-ID: <1992Aug21.015953.10705@decuac.dec.com>
- Sender: news@decuac.dec.com (USENET News System)
- Nntp-Posting-Host: hussar.dco.dec.com
- Organization: Digital Equipment Corporation, Washington ULTRIX Resource Center
- References: <1992Aug19.225010.18306@den.mmc.com> <1992Aug20.042224.11176@decuac.dec.com> <oqme28g@rhyolite.wpd.sgi.com>
- Date: Fri, 21 Aug 1992 01:59:53 GMT
- Lines: 81
-
- vjs@rhyolite.wpd.sgi.com (Vernon Schryver) writes:
-
- >Doesn't Digital offer something you call "NVRAM" which is used with
- >software similar that used with to Legato's PrestoServ? Doesn't ULTRIX
- >while using that "NVRAM" reply to the NFS client before the data is on
- >the disk? Using the reasoning that the "NVRAM" is "stable storage",
- >and so you can call the early reply not really asynchronous?
-
- NVRAM is pretty stable storage. It will survive a power hit
- or something like that. For all intents and purposes I'm willing to
- call that as stable as a disk drive (more stable than some weird caching
- drives can be I suppose).
-
- When running with PrestoServ (we license the technology from
- Legato) you are acking to the client that the data has been copied
- to stable storage, before it's written to disk. You can quibble all
- day about what is or isn't "stable storage" but for damn sure I'd
- rather have my data being written to a battery-backed NVRAM cache
- than into an unprotected buffer cache like can happen on an, *ahem*,
- SGI.
-
- Vern, I've seen you around long enough to know you understand
- these issues. If you're trying to sow fear uncertainty and doubt about
- the effectiveness of NVRAM, why don't you grace us with some arguments
- for how asynch NFS on an SGI is better?
-
- >I have been wondering how much that DEC "NVRAM" costs. Is it standard
- >in all systems?
-
- It's standard on some and it's a memory option on others. The
- memory option is pretty slick, since it's at memory bus speed, not
- on the I/O bus like the Sun PrestoCache board. So you don't chew up
- bus bandwidth.
-
- We get a %300 improvement in NFS ops with Presto on one of
- our servers. Asynch NFS can get you as much as %700, but it assumes
- you're running with a UPS and that no dweezil ever powers your
- CPU off suddenly on you. UPS' are not the only component of a workable
- asynch NFS server.
-
- >When all else is equal, an NFS server with more stable storage is more
- >reliable. However, all else is rarely equal. For example, Ultrix is
- >reputed to had much more serious bugs in its NFS server implementation
- >than any NVRAM or synchronous writes might affect. Actually, "reputed"
- >is too weak. I've seen enough interoperability complaints from SGI
- >customers. I have no doubt DEC will fix those bugs soon, but the point
- >remains.
-
- Gee, Vern, it's so nice of SGI to list our bugs here. Why don't
- you also describe which version numbers of ULTRIX you're talking about?
- I personally know of no major bugs in our NFS implementation in the current
- rev, or I'd be working to get them fixed.
-
- I can handle a little bit of pointed "give and take" on an
- interesting topic, but it's not sporting behavior to post "so and
- so's operating system has *BUGS*" kind of crap unless you care to list
- specific problems with the current rev of the software, and you're
- willing to take your lumps if I start listing any bugs I can think of
- in SGI's. So let's call it quits here while this discussion still
- has a gleam of technical merit.
-
- >It is at best disingenuous to confound UDP checksums which protect
- >against failures on the client, the server, and the network relays,
- >gateways, bridges, and cables in between, with sychronous server writes
- >which protect only against server failures. Both may be desirable by
- >customers, but equating them is wrong.
-
- Equating them is just fine, *IF* they are being sold as an
- option that defaults to being *ON* and *IF* the ramifications of
- asynch NFS are not being explained to the customer. That is what
- I would consider at least mean-spirited if not unethical.
-
- Saying "asynch NFS + UPS is safe" is also not accurate,
- unless you can *GUARANTEE* me that I can't run an application on
- your system that will crash it before it can sync the buffer cache.
- I dunno what "crashme" does on an SGI these days, but asynch
- NFS is a loss, unless you're using it for /tmp or you don't care
- about your data. Or you don't care about your customer's data,
- that is.
-
- mjr.
-