home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!auspex-gw!guy
- From: guy@Auspex.COM (Guy Harris)
- Newsgroups: comp.sys.sun.admin
- Subject: Re: what is prestoserve??
- Message-ID: <15339@auspex-gw.auspex.com>
- Date: 5 Nov 92 05:15:49 GMT
- References: <1992Nov3.042210.22311@afterlife.ncsc.mil> <1992Nov4.212349.19891@sj.ate.slb.com>
- Sender: news@auspex-gw.auspex.com
- Organization: Auspex Systems, Santa Clara
- Lines: 37
- Nntp-Posting-Host: auspex.auspex.com
-
- >I am not sure at what level the write caching is done in the PrestoServe, but
- >in an Auspex file server, their SPIII option which has a similar battery
- >backed write cache feature, even if your NFS write mix is not high, the write
- >cache can have a dramatic impact on total throughput. In the Auspex, the write
- >cache is done at a very low level, caching ALL disk writes.
-
- It's done the same way in Prestoserve (we shamelessly stole the idea,
- although not the implementation, from Legato :-)).
-
- In case some readers don't know why a non-volatile RAM option stuck
- between the file system and disk can speed up stuff other than NFS
- writes, remember that *all* NFS operations are supposed to be done
- synchronously (yes, Vernon, I know; calm down :-)), not just writes, and
- that many NFS operations other than write do writes to the disk -
- create, remove, create directory, rename, etc..
-
- Those operations can also be sped up by a non-volatile RAM option, at
- least in file systems such as the BSD file system, because in those file
- systems, operations that update file system control information are
- often done with synchronous writes, to ensure that they're done in a
- particular order. That's done to minimize the extent to which the
- on-disk data structures can be in an inconsistent state at any point
- during the operation, in order to reduce the chances that "fsck" will
- find a *really* messed-up file system requiring a human to decide what
- to do to fix it.
-
- If "synchronous" means "wait until the data's been written to
- non-volatile RAM" rather than "wait until the data's been written to
- disk", the wait can be shorter. (Yes, if it has to write out a block
- from the non-volatile RAM in order to make room for the new data, it
- won't be; as far as I know - no, I haven't looked at our code lately,
- nor do I remember what the Prestoserve code does - writes aren't
- buffered up until somebody needs the buffer, but are done "in the
- background", so that you don't always stall out waiting for the disk.
- If you're familiar with the way memory subsystems work on, say, modern
- fast microprocessors, think of it as being more like a very large write
- buffer rather than like a write-back cache.)
-