mo@seismo.CSS.GOV (Mike O'Dell) (05/06/87)
In a recent posting here Chris Torek posted some calculations showing that 125 ips 6250 tape drives were "unnecessary" since you couldn't drive them full speed. HORSE HOCKEY!!! The problem with his calculations is that UNIX doesn't do overlapped I/O on raw devices, tape or disk! This means that the REAL TIME for the operation on the tape drive is EVERYTHING! While you may not be able to drive it full speed all the time, you can come pretty close quite often, and with real, high-performance start-stop drives, unlike "streamers", the operations always take just about the same time, ie, not long. I have a Fujitsu streamer on my SUN 3/280, and it streams some large part of the time doing dumps, but when it stops, backs up, and has to get a running start again, it takes forever by comparison. Further, if you every use the tapedrive for anything besides dumps, like really writing tapes full of data, the speed of the drive matters even more since you probably aren't writing the gargantuan blocks that dump uses to get its efficiency. I had a TELEX 8350 (8530??) triple density drive back on a VAX/780 in '82 and have sorely missed it since I left that site. It's sorta like a V8 engine - turbo-charged high-compression 4-cylinders can go quit fast when you get them wound up, but the winding takes a while. A V8 goes fast quickly. (V12's are even better! ;-) ) Nothing beats a REAL tapedrive, i.e., 1600/6250 bpi @ 125 ips start/stop, ideally with 200 ips rewind and fast-forward speeds. Of course these specs are a compromise - until you see 250ips @ 6250 gcr mode, you ain't seen a tape drive. Yours for faster files, and dumps, -Mike le. le.
chris@mimsy.UUCP (05/07/87)
In article <43735@beno.seismo.CSS.GOV> mo@seismo.CSS.GOV (Mike O'Dell) writes: >In a recent posting here Chris Torek posted some calculations showing >that 125 ips 6250 tape drives were "unnecessary" since you couldn't >drive them full speed. >HORSE HOCKEY!!! >The problem with his calculations is that UNIX doesn't do overlapped >I/O on raw devices, tape or disk! *Who* does not? I should have mentioned this. (Indeed, I was going to do so; I cannot now recall why I did not. Perhaps I simply forgot.) 4.3BSD dump uses multiple processes; I use kernel hacks. Either way, we get overlapped I/O. Put the 4.3 dump on your Suns if Sun has not done it for you yet in whatever you are running: it makes an enormous difference. If you use tape for other things, without the asynchronous hacks, yes, tape speed is important. >I have a Fujitsu streamer on my SUN 3/280, and it streams some >large part of the time doing dumps, but when it stops, backs up, >and has to get a running start again, it takes forever by comparison. Streaming drives are killers. Fortunately, a Sun 3/280 with Eagles is a lot faster than a Vax 11/750 with RA81s. >Further, if you every use the tapedrive for anything besides >dumps, like really writing tapes full of data, the speed of the >drive matters even more [...] . (Unless you can convince them to speak asynchronously.) >Nothing beats a REAL tapedrive, i.e., 1600/6250 bpi @ 125 ips >start/stop, ideally with 200 ips rewind and fast-forward speeds. This is `real'? >Of course these specs are a compromise - until you see 250ips @ >6250 gcr mode, you ain't seen a tape drive. Now *this* is more like it! Unfortunately, these drives come with IBM machines and COBOL programmers. :-) Incidentally, if you do image backups: dd if=/dev/rhp4a of=/dev/rmt8 bs=32k (or whatever) from Eagles or faster, you *can* run 125 ips at 6250 bpi. Alas, you then need a spare drive to recover files, unless you like writing *big* programs.... -- In-Real-Life: Chris Torek, Univ of MD Comp Sci Dept (+1 301 454 7690) Domain: chris@mimsy.umd.edu Path: seismo!mimsy!chris
glenn@ll-xn.UUCP (05/07/87)
We have been using two Fujitsu M2436s for about 6 months. They give 200ips at 6250 in start/stop, 500ips rewind. I have verified its 1.25MBps transfer rate using Super Eagles with a VAX 780 and multiple MBAs. This was done, however, with a standalone test program performing direct DMA initiation, i.e., with no in-core memory copying. I truly doubt that such throughput could be sustained with existing dump/restore software on UNIX. BTW, this drive is quite reasonably priced: 25K for one drive with formatter, 15K each for up to 7 additional drives. --Glenn
mangler@cit-vax.Caltech.Edu (System Mangler) (05/09/87)
In article <43735@beno.seismo.CSS.GOV>, mo@seismo.CSS.GOV (Mike O'Dell) writes: > if you every use the tapedrive for anything besides > dumps, like really writing tapes full of data, > the speed of the drive matters even more since > you probably aren't writing the gargantuan blocks > that dump uses to get its efficiency. On the TU80, if you simply buffer enough write data for at least 15 writes and then do them in a tight loop, the TU80 streams in high-speed mode for those writes, and repositions once while you crunch another large chunk of data. You don't need gargantuan writes, although you do need a reasonable-sized buffer (100KB or so). > you can come pretty close quite often [to running it full speed] My SI 9900 disk controller isn't able to keep even a 1600-bpi drive moving at full speed. Whenever dump hits a clump of small files, the controller is 90+% busy, and the tape pauses a lot. What kind of disk controller do you use?!? Don Speck speck@vlsi.caltech.edu {seismo,rutgers,ames}!cit-vax!speck