Subject: Re: seemingly dismal performance of NetBSD-1.1A/sun3 file I/O....
To: Gordon W. Ross <firstname.lastname@example.org>
From: Greg A. Woods <email@example.com>
Date: 04/10/1996 13:46:53
[ On Wed, April 10, 1996 at 11:00:33 (EDT), Gordon W. Ross wrote: ]
> Subject: Re: seemingly dismal performance of NetBSD-1.1A/sun3 file I/O....
> Did you enable DMA and reselect in the SCSI driver?
Yes of course. As I expected though with one SCSI target there's not
much apparent gain between having si_options=3 and si_options=7 (if
memory serves me right).
Here are comparative results from Bonnie for SunOS and NetBSD on
Sun-3's. The first is the SunOS-4.1.1_U1 server, 3/260 w/48MB & the
ST32430N disk. The SunOS NFS test was on a 3/60 w/8MB. The NetBSD NFS
test was on a 3/260 w/16MB. Both NFS tests are to the same SunOS
server and filesystem, the first machine in these results.
-------Sequential Output-------- ---Sequential Input-- --Random--
-Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks---
Machine MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU /sec %CPU
Sun411u1 100 247 94.8 1028 55.4 365 37.4 229 95.1 1294 67.1 52.2 27.4
NFS411u1 100 111 57.0 102 8.6 89 13.7 123 77.3 320 20.5 23.4 19.4
NFS-1.1A 100 77 48.2 109 28.1 84 31.5 117 61.2 267 26.5 20.5 29.2
In this case the machines were not in single user mode, but I avoided
doing anything during the runs so it's just e-mail and such.... I
probably won't be able to do any proper benchmarks in single user mode
again for a few weeks, esp. for the si driver.
In terms of CPU utilization for some aspects of this test, NetBSD and/or
GCC sucks rather badly, esp. if you take into account the fact that the
hardware the NetBSD system runs on is somewhat faster (30% according to
Sun, 4 MIPS vs 3 MIPS or 25MHz vs 20MHz) than the machine the SunOS test
was run on. I've no idea (other than by using lmbench) of how to
compare the memory performance of the two architectures (VME vs SIMMs).
I would guess from this that NetBSD's stdio library needs some work (and
perhaps replacing its guts with sfio would be the right fix), and
otherwise the margin of error probably accounts for the remaining
differences. (Or re-compile NetBSD with the SunOS compiler! ;-)
This difference in overall CPU wastage probably accounts for most of the
differences I see in the slowness of the NetBSD system in booting and
directly running applications.
So, I guess we are back to the si driver, or ffs? Ideas? What should I
try testing next?
I will try to run Bonnie on the disks of both 3/260's in single user
mode as soon as possible.... Hopefully I'll have a 1.1B true -current
kernel for these tests too.
Has anyone been collecting lmbench results for various ports yet?
Greg A. Woods
+1 416 443-1734 VE3TCP robohack!woods
Planix, Inc. <firstname.lastname@example.org>; Secrets of the Weird <email@example.com>