Frank van der Linden wrote:
> All measurements done using hbench-OS, on a P5-200 with 64M of
> memory. I attemped to keep kernel configurations similar, as
> far as possible.

How big has the Pentium cache?
It's quite suspicious the bw performance drops so rapidly
when the code starts to use uvm_loan(). ATM the code maps whole
buffer and transfers it to the reader; transferring by 64KB chunks
might probably help - I'll try. I also need to check what
if hbench doing differently to lmbench to give such wildly different
comparison. You might try how the results would be without
uvm_loan(), if you add a define PIPE_NODIRECT.

It's a good sign that the performance for all >64KB values are mostly
the same for NetBSD version though.

BTW, I see the performance drop with hbench too - the bandwidth
for 128KB transfer size is ~50% of 64KB one on my PIII.

Jaromir
-- 
Jaromir Dolecek <jdolecek@NetBSD.org>      http://www.ics.muni.cz/~dolecek/
NetBSD - just plain best OS! -=<>=- Got spare MCA cards or docs? Hand me them!