Re: NCR benchmarking

Gerard Roudier (
Sat, 6 Jul 1996 22:14:03 +0000 (GMT)

Hi Alexander!

I am very interested in benchmarks.
A bonnie test with a 50 MB file and 64 MB of main memory is something
that I never tried.
In fact you did not need any scsi drivers for this benchmark, a 50 MB ram
disk would be enough.

IMHO, in order to get acceptables results, the file size must be at least 3
times the main memory size.

Can you try again your benchmark with a file size of about 200 MB.

On 6 Jul 1996, Alexander Sanda wrote:

> Now, when we have 2 different drivers for common NCR boards, I was
> interested in performance differences between the "original" linux
> driver, and the ported BSD driver.
> Here are the bonnie results:
> First, the ported FreeBSD driver:
> -------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
> 2.0.3_NC 50 4450 85.0 4375 25.6 1407 17.1 2898 46.1 3849 19.9 66.0 2.9

I get about 3.6 MB/s for input "per char" and 4.3 MB/s for block input
with my 2 years old IBM S12, 100MB file size and 32MB main memory.
I suspect these results to be wrong due to delayed cache flush that
Bonnie cannot see and that the both speed should be the same.

(Due to the size of the file, we must disgard "output" results)

> And now with the linux driver:
> -------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
> 2.0.3_53 50 4656 85.2 4648 24.7 1499 16.7 2895 45.6 4075 17.4 65.6 3.3
> Both tests were run under the same conditions. I did them a short time
> after a clean boot, only one user logged in, same filesystem &
> test-directory. Same hardware, of course :-)
> The linux driver was configured for doing 10 MHz FAST SCSI transfers,
> disconnecting enabled.
> The FreeBSD driver also does FAST SCSI-2, disconnecting and tagged
> queuing.

You were not in the same conditions.
The effects of tagged queue depends on hard disk firmware.
Generally, performances are not affected or are a little increased under
Bonnie. With some HDs, performances are lower with tagged queue.
On the other hand, IMHO, bonnie benchmark is not a serious test for tagged
command queueing.

> System is a ASUS P133 Triton, 64 megs, NoName NCR
53C810. The > benchmarks were done on an IBM DPES-31080 (1 GB)

I will send to you some bonnie benchmarks. One hour will be enough to
get them.