fc5 kernels after 1788 have excessive, kernel overhead

[Anne & Lynn Wheeler]

Andy Burns wrote:
> Even that seems pretty low, I'd have thought you should be able to 
> sustain 10-12MB/s

the example was not taken for being peak tcp/ip ... the example was some
measure of thruput (which happened to include tcp/ip, disk i/o) where
neither the kernel being measure nor the cpu processing should have been
the bottleneck (limiting factor). it turns out that there is some 802.11
involved in this particular operation as well as lots of physical disk
reads on the source and lots of physical disk writes on the receiver.

the issue was that the measured thruput and cpu on 1788 kernel had some
bottleneck outside of the kernel and cpu. the post-1788 kernel thruput
then went to one hundred precent cpu busy ... and the overall measured
thruput, which wasn't otherwise being limited by the system ... degraded
by better than a factor of four ... and the cpu per byte transferred
increased by a factor of over 20 times and the kernel cpu per byte
transferred increased by possibly 100 times.

for some purely random topic drift ... it turns out that slight changes
in the position of the 802.11 antennas can change thruput between
2.6mbytes /sec and 2.4mbytes/sec.