Re: [OFFTOPIC] UltraSparc MMU problems
Richard Gooch (rgooch@atnf.csiro.au)
Wed, 28 Oct 1998 17:08:13 +1100
Hendrik Visage writes:
> David S. Miller wrote:
>
> > Date: Tue, 27 Oct 1998 21:35:16 +1100
> > From: Richard Gooch <rgooch@atnf.csiro.au>
> >
> > David S. Miller writes:
> > > Solaris's TLB miss strategy is pretty hopeless...
> >
> > What does the OS have to do with it? Once the TLB's are set up, why
> > does the OS need to intervene?
> >
> > If the OS uses a stupid scheme to replace a missing TLB entry and uses
> > inefficient tables as backing store for mappings, it will thrash the
> > L2 cache.
> >
> > Note that Solaris and some other svr4 folks swap page tables out to
> > disk when memory is tight... when they get tapped back in you might
> > not get the nice cache patterns you had before...
>
> Note: Richard said it got FASTER after the pageout-pagein operation!!
> Thus, you now have a more efficient set etc. than before!!!
>
> The only thing I think of that would make it FASTER (after the
> pageout/pagein), is that Soalris only page back the needed pages, as needed,
> and thus could be more optimized than before... ie. the spread across memory
> could be less than before the pageout/pagein.
>
> Richard, do try the following (And tell us the results please):
> Repeat your operation as before, but after that:
> - reload the dataset from disk etc. (Thus clearing out the current in memory
> copy with a fresh copy)
> - re-render and time
I get that effect anyway because the dataset is loaded from the
network (via TCP) each time I start the render programme. I don't
actually read from disc.
I can exit and start again and the times stay at around 6.7 seconds.
Regards,
Richard....
-
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/