tomk@sunlamp.UUCP ( Rochester NY) (03/04/87)
Does anyone have a usefull set of benchmarks for testing disk performance under 4.2bsd. The disks would all be attached to the same machine using the same sort of file system so I at least don't have to worry too much about the problems of dealing with different busses and cpus. What I'm looking for (golden fleece) is a reasonably fair set of general disk performance metrics. I'd also like to read any papers available on the subject if any of you have references. -- Tom Kessler Sun Consulting, Rochester NY. sunrock!tomk
henry@utzoo.UUCP (Henry Spencer) (03/11/87)
> What I'm looking for (golden fleece) is a reasonably fair set of general > disk performance metrics... A general comment: the nature of those metrics will depend heavily on what you intend to use the disks for. In particular, the nature of the computing environment will make a big difference. Is it a single-user machine, or a timesharing system / file server? As an example of how these things can affect results, notice that the pre-4.2 benchmarks, demonstrating how wonderful the 4.2 file system was, were all run SINGLE USER. There is a conjecture hereabouts -- not verified by experiment, mind you -- that on a busy multi-user system, the increase in block size accounts for 100% of the performance improvement, with things like cylinder groups contributing absolutely nothing. (What does it matter if all the blocks of your file are on the same cylinder? The heads won't stick around anyway, they've got other customers to service.) -- "We must choose: the stars or Henry Spencer @ U of Toronto Zoology the dust. Which shall it be?" {allegra,ihnp4,decvax,pyramid}!utzoo!henry
mangler@cit-vax.UUCP (03/21/87)
In article <7759@utzoo.UUCP>, henry@utzoo.UUCP (Henry Spencer) writes: > There is a conjecture hereabouts -- not verified > by experiment, mind you -- that on a busy multi-user system, the increase > in block size accounts for 100% of the performance improvement, with things > like cylinder groups contributing absolutely nothing. On the 4.2bsd vaxen that I run, 2/3 of the disk transfers require a seek, and judging from the average transfer time of about 26ms, the seeks are fairly long. (These statistics from a modified disk driver & iostat). When dumping every file on the disk to tape, our vax-780 with its 4K blocksize gets only about twice the disk throughput of the 3b20 sitting next to it doing 1K reads. I'd say Henry is understating his point. Although, even a factor of two can be useful at times... Don Speck speck@vlsi.caltech.edu {seismo,rutgers,ames}!cit-vax!speck
authorplaceholder@gorgo.UUCP.UUCP (03/26/87)
Don Speck writes: >When dumping every file on the disk to tape, our vax-780 with its 4K >blocksize gets only about twice the disk throughput of the 3b20 sitting >next to it doing 1K reads. I would be interested in seeing what the 780 does with 2 and 3K blockss. I also wonder what kind of drives and bus are on the 3b20 and 780... Steve Blasingame (Oklahoma City) bsteve@gorgo.att.com