Subject: seemingly dismal performance of NetBSD-1.1A/sun3 file I/O....
To: None <port-sun3@NetBSD.ORG>
From: Greg A. Woods <woods@most.weird.com>
List: current-users
Date: 04/09/1996 18:52:22
I've been collecting data for some time now in an attempt to narrow this
down, but I have not made it much further than my first tests.
It appears that NetBSD-1.1A (~960313), running on a Sun-3/260 with the
si SCSI disk driver, is much slower than SunOS-4.1.1_U1 on the same
hardware. It may also be true that NetBSD-1.1A is slower when running
diskless.
I have two pairs of comparison systems. The first are disk-full and
nearly identical:
Sun 3/260
48 MB RAM (2x8+2x16)
si scsi controller
sd2: <ST32430N 2.4Gb cyl 3992 alt 2 hd 9 sec 116>
The SunOS machine also includes some other disks, but the test results I
include at the end of this message were obtained in single user mode on
both systems with the same kind of disk. Both had cgtwo consoles at the
time of the tests, though the SunOS machine has since gone headless.
sd0: <Quantum PD 1050S cyl 2444 alt 2 hd 12 sec 70>
sd4: <Maxtor XT-3280S cyl 1222 alt 2 hd 15 sec 25>
The other two diskless machines are not identical but similar enough
that I have ample experience with what to expect on them:
SunOS-4.1.1_U1 + X11R5 NetBSD-1.1A/sun3 + X11R6
Sun 3/60 Sun 3/260
8 MB RAM (SIMM) 16 MB RAM (2x8)
bwtwo0: 1600x1280 mono cgtwo0: 1152x900
Given the performance I've seen on both these pairs of machines, I'm
suspecting the problem is more in the filesystem than in the SCSI
interface, though both may be contributing (the overhad of the extra
data in the new filesystem may slow the overall system down, but should
not have slowed down the direct disk I/O measurements). The diskless
NetBSD station, running X11R6 may also be suffering from the bloat in
that version of X11. I had run an 8 MB RAM 3/260 diskless with SunOS
and X11R5 and found it to be slightly faster than the 3/60 with the same
software. To now find any file I/O operations running at what seems to
be half the speed or less under NetBSD on the 3/260 that now has twice
the memory is very disappointing, to say the least.
The diskless NetBSD boot takes about twice as long as the SunOS boot
(both booting from the same SunOS-4.1.1_U1 server at separate times).
Starting X11+xdm+xclock+xconsole+xearth takes almost three times as long.
Once it's up and running though, the X11 server itself seems responsive
enough, and clients running on other hosts are quite OK of course.
I suspect some of the overhead is in the larger data structures in the new
Now for the hard figures on the Seagate ST32430N drives, in single user
mode, first from the SunOS system, then the NetBSD system:
ttyp3:<woods@most> $ ./iozone auto
IOZONE: Performance Test of Sequential File I/O -- V2.01 (10/21/94)
By Bill Norcott
Operating System: SunOS -- using fsync()
IOZONE: auto-test mode
MB reclen bytes/sec written bytes/sec read
1 512 582442 782432
1 1024 770853 1379652
1 2048 953077 1872360
1 4096 1027854 2183855
1 8192 1048418 3084056
2 512 563662 748860
2 1024 748859 1294316
2 2048 979835 1191355
2 4096 859381 1519339
2 8192 476556 3494501
4 512 414394 537646
4 1024 533541 896086
4 2048 611320 979813
4 4096 1115325 1530528
4 8192 1022846 1164887
8 512 415214 731882
8 1024 801851 1397887
8 2048 1072544 1327105
8 4096 942395 1481842
8 8192 1130355 2932609
16 512 535587 683564
16 1024 775167 1060347
16 2048 1038036 1492439
16 4096 1094969 1139607
16 8192 1099267 1923732
Completed series of tests
ttyp2:<woods@insanity> $ iozone auto
IOZONE: Performance Test of Sequential File I/O -- V2.01 (10/21/94)
By Bill Norcott
Operating System: POSIX 1003.1-1990
IOZONE: auto-test mode
MB reclen bytes/sec written bytes/sec read
1 512 270251 1038194
1 1024 298739 1664406
1 2048 340446 2330168
1 4096 346064 2833989
1 8192 794375 3177503
2 512 270600 791378
2 1024 309771 1651300
2 2048 334474 2330168
2 4096 324636 2833989
2 8192 816012 3226387
4 512 272004 656385
4 1024 316312 739736
4 2048 335812 817603
4 4096 341277 820803
4 8192 812849 791378
8 512 272445 647269
8 1024 315124 710297
8 2048 337162 767484
8 4096 338113 791378
8 8192 795883 792874
16 512 272401 645029
16 1024 315065 722532
16 2048 336959 765733
16 4096 330455 797775
16 8192 788773 795883
Completed series of tests
Note that a "dd if=/dev/mem of=/dev/null" is about 4 times faster (if I
remember right from my last posting on this topic) on NetBSD than SunOS.
All of this seems very strange to me, since NetBSD-1.0A on SPARC seemed
much faster than SunOS-4.1.3.
--
Greg A. Woods
+1 416 443-1734 VE3TCP robohack!woods
Planix, Inc. <woods@planix.com>; Secrets of the Weird <woods@weird.com>