Subject: Re: Port benchmarks
To: Manuel Bouyer <bouyer@antioche.lip6.fr>
From: Sean Doran <smd@ebone.net>
List: port-alpha
Date: 04/06/2000 16:57:37
Manuel Bouyer <bouyer@antioche.lip6.fr> writes:
> Another result, ds20-1/NetBSD 1.4X is a system (kernel + userland + lmbench)
> compiled '-O3 -mcpu=21264 -mmemory-latency=main -fsched-interblock
> -fsched-spec -fsched -spec-load -fbranch-count-reg' (this is basically what
> simon gave me :)
> So, a small gain for most latencies, nothing for bandwidth. Still far from OSF1
> in many aspects :(
Our egcs is nowhere near OSF1's cc. :(
However, I would be interested in seeing if you get any
visible results from playing with -mmemory-latency=
settings, given that you have fast & biggish caches.
Look for "alpha_mlat_string" in /usr/src/gnu/dist/gcc/config/alpha/alpha.c
In /usr/src/gnu/dist/gcc/invoke.texi there is a brief
blurb about what it's all about - note that the value is
described as "[a] decimal number representing clock
cycles", and is about as good a description as you will
find. :( I imagine that it's the nominal clock cycles
needed to fetch a word of memory on the system one expects
to run the binary upon. [i.e., if you blow your very fast
cache alot, you want to use a higher value]
Sean.