John Fremlin wrote:
> 
> Single copy:
> 
> > Pipe bandwidth: 194.09 MB/sec
> 
> About 59.5 (+-0.6) Mb/s :-(
>
You are testing the L2 cache:
K6-2: 0 kB
PPro: 256 or 512 kB Full speed

I know that on Linux bw_pipe with the standard 64 kB buffers fits into
the 256 kB cache of my Pentium III. Total bw_pipe bandwith with 2
Pentium III's : 2.6 GB with simple PC100 memory ;-)

But you really shouldn't use bw_pipe to compare OS'es. What you are
measuring is

* the efficiency of the memcpy implementation. Does the K6-2 support
'movntq'? That might give you a 30% boost with bw_pipe [but is probably
slower in real world cases]
* the scheduler overhead. Check the number of context switches for one
bw_pipe run.
* the overhead of the memory management layer. Linux has a slight
advantage since it doesn't need uvm for i386 with < 900 MB ram, and
above 900 MB ram it uses a 2/4 MB cache of the last mappings.

--
	Manfred