On Sat, Jul 29, 2000 at 06:23:48PM -0700, Kanoj Sarcar wrote:
> And after the recent tlbflush optimizations:
>
> SMP
> ---
> 771.97user 55.22system 4:24.73elapsed 312%CPU (0avgtext+0avgdata
> 0maxresident)k
> 0inputs+0outputs (508064major+1831879minor)pagefaults 0swaps
>
> I am still trying to figure out the DISCONTIGMEM problem, but at least
> it seems that on specific memory configurations, not ignoring the 128Mb
> memory on the 7th bank seems to create problems. Go figure.
I'm rewriting some important routines like memcpy, memmove and ip checksums.
As a result of the first crude test implementation the results of
microbenchmarks like lmbench have increased significantly, sometimes beyond
factor of two.
A analysis of lmbench also shows that some of our number degrade much worse
than IRIX's for large datasets; the graphs seem to indicate that the lack
of cache colouring is the cause.
Ralf
|