douglis@ginger.Berkeley.EDU (Fred Douglis) (10/27/87)
[This is in response to a request for information on optical disks and Sprite; I thought it might be of interest to the general community, and I'd be interested in comments. - FD (as opposed to DL)] Sorry, I haven't written anything on the Sprite archiving system yet -- I'm only just starting to shift the emphasis of my research toward optical disks. I can't tell you too much about disk availability, but I know a little. Over the summer at DEC WRL I used Maxtor 800-S drives, 400MB/side, double-sided, cost $1500 each directly from the manufacturer. Here at UCB, we apparently couldn't get them from Maxtor, so we got them from a company called "Corel" in Ottawa -- don't know the cost. Same drives, though. We just got them last week and have not yet hooked them into Sprite. I understand people elsewhere are using or have ordered drives from Optimem, as well as Sony and other Japanese firms. Don't know the cost. It seems clear that in the long run, for a centralized archive service you need a jukebox with many disks rather than a small number of scattered small drives; however, Hugh Lauer (of Apollo) mentioned that he thinks it'll be a while before jukeboxes work well enough to use and thinks instead that people should get large numbers of small optical disk drives, such as one per workstation plus a few more for general use. Cosmos, with its immmutable objects, would be a natural place to use optical disks, although I worry a bit about storage space: if *every* past version of *every* object is available, that's a lot of space. Lot's of people talk about writing out to an optical disk only if a file survives past a certain lifetime, and that may be the way we'll go in Sprite -- however, it means that there may be gaps in the versions in the archive. Another issue is to distinguish which files have every version written, and to write every version of things like C files but not to write out the binaries. Do you know what you'll be doing about this? - Fred - douglis@ginger.Berkeley.EDU ucbvax!douglis
preston@felix.UUCP (Preston Bannister) (10/28/87)
In article <4177@sdcsvax.UCSD.EDU> douglis@ginger.Berkeley.EDU (Fred Douglis) writes: >It seems clear that in the long run, for a centralized archive >service you need a jukebox with many disks rather than a small number >of scattered small drives; however, Hugh Lauer (of Apollo) mentioned >that he thinks it'll be a while before jukeboxes work well enough to >use and thinks instead that people should get large numbers of small >optical disk drives, such as one per workstation plus a few more for >general use. We build and sell a _lot_ of optical disk jukeboxes as a part of our system. They work quite well. As to whether to use a central jukebox or a individual drives, you should keep in mind the fact that optical disks are not as of yet an especially fast medium. In our system the optical disks are accessed through a server. The server is responsible for satisfying read and write requests in the most efficient possible manner. The server has a hard disk for staging data to be written to optical disk and caching data read from optical disk. Writing data to the server is, in general, faster than writing to optical disk, as the server will first write the data to magnetic disk (which is faster). Once the data is written to the server's hard disk, the client can go about it's business. The server will actually write the data to optical disk at some later point in time. There is a place for individual optical disk drives. If you are continuously writing large volumes of data to optical disk, it may be more effective to have a disk attached to the machine generating the data. The central jukebox is a win when a large number of people wish to share the same data. -- Preston L. Bannister USENET : ucbvax!trwrb!felix!preston BIX : plb CompuServe : 71350,3505 GEnie : p.bannister