Subject: Re: seemingly dismal performance of NetBSD-1.1A/sun3 file I/O....
To: Curt Sampson <curt@portal.ca>
From: Michael L. VanLoon -- HeadCandy.com <michaelv@HeadCandy.com>
List: current-users
Date: 04/09/1996 23:13:32
>On Tue, 9 Apr 1996, Greg A. Woods wrote:
>> 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....
>> 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...

>This is just a thought, but has there been any attempt to analyse
>the CPU usage of filesystem operations and perhaps optimise them?
>I've got a 486/100 with an NCR controller and a P90 with a Buslogic
>controller, both with an HP Surestore 5400 RPM disk drive (among
>others). Using iozone, the P90 gets about 4MB/sec on the HP drive,
>the 486/100 gets about 1.5 MB/sec. The results on the 486/100 vary
>depending on whether the filesystem is small or large, and how full
>it is. The P90 doesn't. This leads me to suspect that something in
>the CPU may be a limiting factor here.

Odd.  I assume your NCR is PCI.  I get better figures from less
machine.  I'm able to get right around 3MB/s on a Quantum TrailBlazer
850, and around 4MB/s on an HP SureStore 3323 (1GB) with iozone, on
the 16MB test.  And, even a little better than that, ~4.2MB/s on a ccd
striped across both of those drives.  I'm probably getting close to
CPU saturation at that point, with the ccd.

The machine is a 486DX2/80, EISA bus, BusLogic BT747s EISA controller,
512K write-back cache, and 24MB RAM.

I do notice that fuller filesystems tend to drop performance a little
bit for benchmarks, however.  These test were run on filesystems less
than 50% full, with 2K/16K fragment/block size.

Bonnie shows CPU usage to be in the mid to upper 90% range for the
per-character transfers, as expected.  And, in the teens for block
transfers on the single drives, and 30%-50% for my ccd partition.
The three-daemon random-seek test putting CPU usage between 2.8% and
5.1%.

I would suggest trying Bonnie out on your systems, too, to see how
something other than linear reads and writes performs on different
platforms.

For what it's worth...

-----------------------------------------------------------------------------
  Michael L. VanLoon                                 michaelv@HeadCandy.com
        --<  Free your mind and your machine -- NetBSD free un*x  >--
    NetBSD working ports: 386+PC, Mac 68k, Amiga, Atari 68k, HP300, Sun3,
               Sun4/4c/4m, DEC MIPS, DEC Alpha, PC532, VAX...
    NetBSD ports in progress: PICA, others...
-----------------------------------------------------------------------------