Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!mnetor!uunet!husc6!cmcl2!beta!hc!ames!sdcsvax!darrell
From: douglis@sloth.Berkeley.EDU (Fred Douglis)
Newsgroups: comp.os.research
Subject: Re: Optical disk questions
Message-ID: <4134@sdcsvax.UCSD.EDU>
Date: Mon, 19-Oct-87 19:58:09 EDT
Article-I.D.: sdcsvax.4134
Posted: Mon Oct 19 19:58:09 1987
Date-Received: Tue, 20-Oct-87 23:49:12 EDT
Sender: darrell@sdcsvax.UCSD.EDU
Organization: U.C. Berkeley
Lines: 54
Approved: mod-os@sdcsvax.uucp


I'd like to respond to Lawrence Lerner's note about optical disks.
I'm a grad student at U.C. Berkeley working on the Sprite project; I
am currently starting research on the use of optical disks for
archiving and file migration.

There are a number of write-once optical disks already available on
the market, and write-many disks will be available in the next few
years.  I have had some experience with the Maxtor 800S drive:
400MB/side, double-sided, SCSI-based.  Last summer at DEC WRL, I
implemented a primitive archive service using the Maxtor drive -- we
measured transfer rates on the order of 30KB/sec writing and 70KB/sec
reading, with no measurements of the latency at the time I left.  I
believe that the 12" drives have much better transfer rates than the 5
1/4" drives; also, the device driver was not at all optimized, so the
rates I measured were pessimistic.  Still, there's a big difference
between the performance of a magnetic disk and that of an optical
disk, just like there's a big difference between memory and magnetic
disks.  Treat the optical disks like another level in the cache
hierarchy, and everything's just fine.  

We asked ourselves what one could do with vast amounts of write-once
storage, and we decided to try to use optical disks to store old
versions of files, so that people could turn the clock back to see the
file system as of an earlier point in time.  The same mechanism could
be used to remove unused files from magnetic disks and restore them
(in a "short" time) from optical disks if they are later referenced.
Large memories and magnetic disk caches would be used to keep the I/O
to the optical disk to a minimum.

A number of issues come up with respect to automatic archival of old
files.  We have to decide how to deal with log files (we wouldn't want
to rewrite  several megabytes every time a few bytes are appended),
and we have to perform measurements to decide if it's feasible to
store all versions of all files or to exclude large, recreatable files
such as executables.  The file system would probably have to be
modified to tag files with attributes (temporary, append-only, etc.).
We must decide on data structures to facilitate efficient access to
old versions of files: given a time, what version of each file (if
any) existed at that point?

The research I'd like to do is pretty preliminary at this point, and
I'd welcome feedback from anyone who would like to comment.  Right
now, I'm reviewing previous work in the area (e.g., Swallow (MIT), and
file migration work at the University of Illinois).  I'd appreciate
any pointers to other relevant work, or any other suggestions.  Please
direct them to:

			douglis@ginger.Berkeley.EDU
			...!ucbvax!douglis

- Fred -

- Fred -