On Mon, 17 Jan 2005 22:43:38 -0500
Allen Briggs <briggs@netbsd.org> wrote:

> I'm kind of surprised to see a 5-10% drop off in ftp performance for a
> 10Mbps ethernet interface.

I've done some initial benchmarking, which has revealed additional
diagnostic information. There are clearly still some problems with mc*,
and an additional few percent performance (maybe more) to be squeezed
out of mc*.

With five runs of putting a 7M file via ftp, the MP kernel I'm using
will average .99MB/sec. The SP kernel compiled with the same source
code will average 1.04MB/sec (5% drop-off). With five runs getting the
same file, the MP kernel will average .99MB/sec. The SP tends to be
unable to successfully get the same file, as I get repeated mc0: babble
messages and then the chip resets repeatedly. When I was testing my
patch on the receive buffers, I was unable to test getting via ftp as
the kernel I was using was an MD and it was out of space to write the
file. This is why I missed seeing there were still some problems. The
above benchmarks were run on disk booted kernels. The SP kernel is quite
useable for ssh even with the above problems.

I find that the SP kernel has less variation in its throughput numbers.
This was also present in the ping numbers I posted, as the deviation was
noticeably less than with MP.

While trying to get the file with SP, I'm seeing performance around
780KB/sec, so this can obviously improve once the errors are eliminated.
I will be examining the remaining problems with mc* and post any patches
that I come up with. 

With both kernels, I get mc0: babble messages, but it tends to be
overwhelming and terminal with SP more than it is with MP. If I can
isolate and eliminate this problem, I believe throughput will improve
above the listed values above.

tim