Andrea Arcangeli wrote:
>
> On Wed, 21 Jun 2000, Tom Leete wrote:
>
> >This is the summary report of four lmbench runs each. I can
> >provide the detailed report if there is interest. It appears
>
> Many thanks for the feedback. However I have to say that things like
> Processor, Context switching Local communication latencies, Local
> communication bandwith and memory latencies in nanoseconds can't be much
> differnt. You should do heavy allocation of memory to exploit the
> difference between the two kernels.
Yes, these figures are for very light loading -- fresh boot,
text mode, mc running. They say more about L1 and L2. Disk
i/o is tested, but swap is not exercised.
> >From lmbench there's some test that bench how fast the machine
> reads/writes from/to memory. If you cause such test to run into swap (so
> if you tell lmbench you have more memory than the amount of physical RAM)
> it will also benchmark how the machine reacts to low memory scenario for
> example. Doing that you could get more different numbers I think. And btw
Agreed. I have a few runs for the kernels with two stages of
memory loading from X, gnome soup, and netscape. Incomplete
yet.
> make sure to apply the patch that I posted to the list against ac22-class
> before bechmarking ;). That patch backout one bugfix and something
> introduced into 2.3.99-pre that is going to hurt performance. I'll attempt
> to fix it right (in a way that can't hurt) soon.
Ok, thanks. I'll hold off on the loaded runs till then.
One statistic that lmbench keeps is the number of page
faults. For carefully prepared loads (lots of history
dependence here), we can look directly at the accuracy of
different swap strategies.
I should look into calibrating these in terms of
/proc/meminfo
Regards,
Tom
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Fri Jun 23 2000 - 21:00:23 EST