Manuel,

Do you have OSF that you can install on this machine so as you can
verify my original DS20 results?  It seems that the mmap latency
figures on NetBSD are slow for some reason, and possibly the filesystem
benchmarks if the ds20 I had access to had some sort of hardware
assistance (I got some pretty impressive figures on a 40MHz MIPS R3000
DEC box running Ultrix that had prestoserve).  It would be interesting
to see a disassembly of the bcopy routine out of libc - I wonder if they
have a different bcopy for each processor revision or something wierd
like that...

I just had a peek at our bcopy - there's appears to be no loop unrolling
at all (look at the bcopy_samealign_lp).  What's the typically loop
overhead on an alpha CPU?

Simon.