mcneild@mcneild.Software.Mitel.COM (Doug McNeil) (11/11/90)
First thanks to those who helped with my ULIMIT problem... As the subject says I'm interested in comparing the performance of the various flavours of UNIX running on 386's. Do porters usually do significant performance tuning? ESIX, for example, advertises performance specs which are significantly better than SCO 3.2. Driver differences I understand, but things like _system_call_overhead_? Do these people actually tweak the AT&T sources (and should this make us happy, apprehensive?)? (My interest stems from misspent youth, when I used to toil untold hours writing in assembler, trying to extract more oomph from PDP-11's ;-) Thanks -- -----------------------------------------------------------------------------> "Smokin" Doug McNeil | ...!uunet!mitel!spock!mcneild Mitel Corporation, Kanata, Ontario | (613)592-2122 "I know a lot of fancy dancers.." - Cat Stevens |
davidsen@sixhub.UUCP (Wm E. Davidsen Jr) (11/14/90)
I have no doubt that the results Everex quotes so frequently are true,
in that they really ran benchmarks of some nature and got the results
they show. I have not verified that this reflects the number I would get
from another benchmark, or that either is a good predictor of how well
the system would perform under normal load.
When I benchmark I do measure system call overhead, but only for a few
calls. I use getuid(), since that is unlikely to involve any i/o in a
rational system.
I also measure file system throughput for single large files. I write
a single file of size 6MB or size of physical memory, whichever is
larger. I then read it back in. I measure the time it takes and get a
write speed and read speed. I find this useful to me, because it
reflects performance on the largest file I am like to copy on a regular
basis on that machine. I realize that caching will help the performance,
but it will in real application, too, so that's fair.
Using this test of disk performance I find that SCO Xenix has better
performance than any of the V.3 or V.4 variants I've tried.
Test conditions: unloaded system, multiuser mode, typically 4MB to 8MB
memory. Write performed by a C program looping writing 1k blocks. Time
from the time system call.
You are the sole judge of how well the results of this test apply to
any productive work you may run on the machine.
--
bill davidsen - davidsen@sixhub.uucp (uunet!crdgw1!sixhub!davidsen)
sysop *IX BBS and Public Access UNIX
moderator of comp.binaries.ibm.pc and 80386 mailing list
"Stupidity, like virtue, is its own reward" -mefangchin@elaine41.stanford.edu (Chin Fang) (11/16/90)
bill> Using this test of disk performance I find that SCO Xenix has better bill>performance than any of the V.3 or V.4 variants I've tried. bill> bill> Test conditions: unloaded system, multiuser mode, typically 4MB to 8MB bill>memory. Write performed by a C program looping writing 1k blocks. Time bill>from the time system call. bill> bill> You are the sole judge of how well the results of this test apply to bill>any productive work you may run on the machine. bill>-- bill>bill davidsen - davidsen@sixhub.uucp (uunet!crdgw1!sixhub!davidsen) Bill's test is very similar to a comprehensive test of sco xenix 386, micropot 386/at unix, enix(now esix) rev.a, and a few others conducted by MIPS magazine (now Personal Workstation) about two years ago. MIPS testers used very similar machine configurations to Bill's. And most interestingly, their conclusions are also close to Bill's given above. (MIPS obviously didn't have R4 available for test, however) I still remember that one MIPS tester mentioned that in terms of memory requirement, SCO 386 xenix won hands down. Over years of refinement(?), SCO indeed reduced the kernal size of 386 than any of UNIX V5 R3.2. People interested in what MIPS did can go to library and check it out. MIPS is somewhat a new magazine started barely three years ago(?). Chin Fang Mechanical Engineering Department Stanford University [Rockwell International, Rocketdyne Div.] fangchin@portia.stanford.edu fang@rocket.cadcam.rok.com
david@talgras.UUCP (David Hoopes) (11/16/90)
In article <2285@sixhub.UUCP> davidsen@sixhub.UUCP (bill davidsen) writes: > > Using this test of disk performance I find that SCO Xenix has better >performance than any of the V.3 or V.4 variants I've tried. > I did some testing awhile back comparing file access speed on SCO Xenix 2.3.2 and SCO Unix 3.2 My tests involved reading 10 1meg files, 200 small files in one directory, and about 50 small files each in its own directory. Both tests where performed on the same 33mhz 386 and the same type of hard drives. I was trying to find out how fast I could read these files off of the filesytem. I used the following command and timed it with a stop watch: tar cvf /dev/null filenames With a brand new filesystem: Xenix averaged 13 meg/min Unix averaged 23 meg/min I ran a shell script that backed up the files to tape, removed the files, restored the files, and then tested them (I was testing the tape driver), and then repeated. I let it run for 24 hours and then repeated the time test. Xenix averaged 3 meg/min Unix averaged 22 meg/min Unix was faster to start with and had very little degredation from the repeated writing and removing the files. -- ------------------------------------------------------------------------ David Hoopes Tallgras Technologies Inc. uunet!talgras!david 11100 W 82nd St. Voice: (913) 492-6002 x323 Lenexa, Ks 66214