On Sun, Apr 09, 2000 at 12:14:41AM +1000, Simon Burge wrote:
> > Simon asked for a Linux/NetBSD lmbench comparison on my Cobalt box.
> > 
> > The lack of a proper microtime() on Cobalt right now made lmbench act
> > a little funny sometimes (and made the runs take a long time),
> 
> Is this the "taking 297 seconds to calculate CPU speed" and so on?  The

Yah, and also 'gettimeofday() takes 0 usecs' and such.

> same problem exists on Alphas as well - I always thought it was neat the
> my 60MHz R4400 took _way_ shorter than my 500MHz 21164 for an lmbench
> run but never figured out why the timing goop was so different.

I haven't checked the algorithm, but lmbench seems to really like/want
better than 100 Hz gettimeofday() resolution.

I intend to hack up a little MIPS3 cycle counter based microtime() and
rerun the test, though I am not enough of a time geek to know how exactly
to make it stabilize itself.

> > so not
> > all numbers may be dependable, but at least some of the results seem
> > to indicate that we are somewhat behind Linux on MIPS at the moment.
> 
> The numbers don't seem too bad compared with the DEC 60MHz R4400 boxes,
> except the the mmap latency is ~15 times faster.  I've been seeing quite
> shocking figures on both pmax and alpha - that figure is amazingly good.

I am rather suspicious of some of those mmap latency figures.


-- 
Soren