How fast are your disks?
Steve Knodle
slk at clutx.clarkson.edu
Tue Apr 5 07:09:33 AEST 1988
Concerning the following discussion:
>From: Tim Bray <tbray at watsol.waterloo.EDU>
>Subject: How fast are your disks?
>Date: 1 Apr 88 19:09:18 GMT
...
>In a recent meeting we were analyzing the performance of this application that
>is rather I/O bound - in particular, it performs a lot of very random accesses
>here and there in large (> 100 Mb) files. Somebody said "Now, we'll assume
>that Unix can do a maximum of 30 disk I/O's a second". Somebody else remarked
>that that figure had been remarkably constant for quite some time. Somebody
>else proposed that it was a fundamental law of Computer Science. (Of course,
>we are poor peons restricted to the use of Vaxes and Suns).
>
>Anyhow - Presumably there are other people out there limited by this
>particular bottleneck. Are there reasonably-priced unix systems out there
>that do better? Are there a set of benchmarks which reliably characterize
>system performance in this area?
>From: Ron Natalie <ron at topaz.rutgers.EDU>
>Subject: Re: How fast are your disks?
>Date: 3 Apr 88 00:04:52 GMT
>Keywords: Disk I/O throughput
>To: unix-wizards at brl-sem.arpa
>
>Not only is it not a constant, it's not even true. The sad fact
>is most disk controllers for minis/micros are pretty horendous.
>Sun's unfortunate use of the Xylogics 450/451 is a prime example.
>Anyway, with decent controllers (or multiple controllers) there is
>no reason why the figure 30 can't be exceeded and is on decent Unix
>systems.
>
>-Ron
Let me offer, as a point of reference, extracts from simultaneous (vm/io)stat
performance logs for our Gould Powernode 9080, which performs very gracefully
under severe I/O load, I feel. The logs were taken during the End-of-Semester
Crunch, and are being used to substantiate my request that the machine
be upgraded from 8 Meg of memory to 16 Meg. (The relatively small
amount of memory explains the severe paging rate below.)
Our 9080 serves as a campus timesharing host serving the engineering
and scientific faculty and student computing. The job mix is a combination
of large background jobs (finite element jobs and fluid-flow simulations)
with editing, compiling, and electronic mail as foreground tasks. There
were about 25 - 35 users logged on during the period below. Response time
was decently good. Few users noticed that the load average had drifted up
until we reached the threshhold at which sendmail stops delivering and only
queues. They started wondering why e-mail wasn't being delivered! So I think
that significant I/O capacity still remained.
The main contribution to its I/O performance is the combination of Gould's
good High Speed Disk Processor and the CDC 858 MB disk (our device "dk0").
This pair is reportedly capable of reading and buffering an entire cylinder in a
single I/O operation. For large sequential files, this and the Berkeley
Fast File System cylinder grouping gives a big win. I suspect this explains
the very last extract below where transfers/sec hit 60.
The complete machine configuration is: 2 CPU's, 3 disks on 3 controllers
(1 HSDP and CDC 858 (18 ms) and two older CDC 650's (25 ms) on UDP controllers).
Swapping is distributed on all three disks. One third of the users come in
via ethernet, the rest via ttys.
The logs were taken with a script that alternated "date" with "vmstat(iostat) 180 20"
The second CPU isn't included in cpu usage under this display format, and
spends its time almost exclusively in user mode.
---------------------- vmstat log -------------------------------
procs memory page dk faults cpu
r b w avm fre re at pi po fr de sr d0 d1 d2 d3 in sy cs us sy id
.......
2015 0 17992 152 8 0 296 80 224 0 101 21 6 16 0 902216 231 67 32 1
24 0 0 17552 360 6 0 104 72 104 0 69 35 7 16 0 772015 121 61 37 1
1718 2 19272 48 9 0 664 120 280 192 133 33 8 13 0 120 758 376 58 41 1
24 1 1 17784 104 11 0 400 176 280 0 116 22 6 11 0 881293 147 58 41 1
25 2 0 19144 312 10 0 320 136 200 0 98 19 6 14 0 117 492 196 62 38 1
21 6 0 17080 208 5 0 392 160 248 0 87 27 8 15 0 751885 128 65 34 1
20 5 0 16464 64 7 0 432 168 264 0 130 17 5 12 0 881113 178 61 38 1
25 1 0 15728 416 12 0 136 80 144 0 70 23 6 13 0 392390 55 61 38 1
21 5 0 16808 128 9 0 504 128 280 0 131 17 6 13 0 70 505 139 70 29 1
22 5 0 16000 152 5 0 232 80 160 0 124 25 7 16 0 76 484 119 76 23 1
18 5 6 19192 80 7 0 320 88 184 416 91 25 8 17 0 84 252 143 74 25 1
17 5 6 19376 88 6 0 464 208 280 256 133 36 12 18 0 116 512 176 66 33 1
22 5 0 15448 160 9 0 280 88 160 0 111 30 7 13 0 84 395 123 74 25 1
22 4 5 18000 56 8 0 448 136 216 216 99 33 11 21 0 98 499 170 64 35 1
---------------------------- iostat log -------------------------------
tty dk0 dk1 dk2 dk3 cpu
tin tout bps tps msps bps tps msps bps tps msps bps tps msps us ni sy id
...
16 467 200 18 16.2 77 5 15.4 182 12 16.4 0 0 0.0 38 25 36 1
15 379 230 20 16.8 93 6 15.6 217 13 16.2 0 0 0.0 37 38 24 1
19 473 279 26 17.7 111 7 15.6 276 17 17.6 0 0 0.0 23 51 25 1
18 606 318 30 17.3 141 10 18.7 265 17 17.3 0 0 0.0 34 38 28 1
16 353 333 33 18.5 163 11 17.5 248 16 17.4 0 0 0.0 43 25 30 1
19 493 302 30 17.3 117 8 17.2 234 16 16.7 0 0 0.0 27 42 30 1
19 458 344 35 17.8 170 11 17.4 294 21 17.8 0 0 0.0 39 26 33 1
19 555 340 35 18.2 187 13 17.9 282 20 17.5 0 0 0.0 44 20 34 2
13 552 388 39 17.5 160 11 17.1 221 15 17.2 0 0 0.0 41 25 32 1
12 443 340 30 17.5 131 9 18.6 228 15 17.1 0 0 0.0 41 33 24 1
15 301 348 30 17.7 133 9 18.6 245 17 17.4 0 0 0.0 32 40 26 2
13 355 426 41 17.2 138 9 18.0 231 17 16.9 0 0 0.0 44 22 33 1
24 915 394 35 17.8 151 11 19.4 238 16 17.5 0 0 0.0 40 27 32 1
-------------------------- iostat extract ---------------------------
Sat Dec 12 07:29:18 EST 1987
tty dk0 dk1 dk2 dk3 cpu
tin tout bps tps msps bps tps msps bps tps msps bps tps msps us ni sy id
...
0 3 201 38 13.6 14 1 15.9 10 1 13.5 0 0 0.0 19 64 16 1
0 3 326 61 10.1 6 0 15.6 3 0 13.1 0 0 0.0 18 66 15 1
0 2 138 24 13.4 14 1 13.8 16 1 13.9 0 0 0.0 55 29 15 0
More information about the Comp.unix.wizards
mailing list