[linux-mm-cc] Announce: ccache release 0.1

Nitin Gupta nitingupta910 at gmail.com
Fri Feb 22 03:38:30 EST 2008


On Fri, Feb 22, 2008 at 1:56 PM, John McCabe-Dansted <gmatht at gmail.com> wrote:
> On Tue, Feb 19, 2008 at 11:04 PM, Nitin Gupta <nitingupta910 at gmail.com> wrote:
>  >  > > > BTW, why is the default 10% of mem?
>  >  >
>
> >  > Perhaps 100% (or maybe 50%) would be a more sensible default? For me
>  >  > 66% makes a huge difference to the Hardy liveCD performance. 10% makes
>  >  > a difference but 50%+ goes from "ls /" taking 10s to snappy
>  >  > performance even on large applications like Firefox.
>  >  >
>  >
>  >  I think this depends a lot on kind of workload and system. For e.g:
>  >  - On desktops, retaining too many anonymous pages at cost of
>  >  continuously losing page-cache (filesystem-backed) pages can hurt
>  >  performance for workload that repeatedly access same file(s).
>  >  - On embedded systems, too much de/compression will drain all battery.
>  >  and so on...
>
>  I was thinking that for the LiveCD case the primary performance hit
>  would be seeking on the CD drive. Hence to maximize memory available
>  for caching the CD, we should
>    1) Use 100% mem for compcache.

Agree.

>    2) Set compcaches priority to be low, so that hdd swap will be used
>  first if it exists.
>  Does this seem reasonable?

Setting compcache prio < hdd swap will not be good. After all,
compcache can potentially avoid swapping to slow hard-disks improving
system performance significantly. If compcache is not sufficient to
store all pages then only we should let it go to lower prio swaps.


>
>  >  Unfortunately none of these messages suggest why crash happened.
>  >  If you can send entire log, that will probably be more useful.
>
>  I now tested this with QEMU and -smp 2. Still cannot reproduce the
>  bug. QEMU doesn't seem to use both host CPUs so I presume it doesn't
>  interleave instructions nearly as much as real SMP would.
>
>  I'll test this some more on my real hardware (crosses fingers)
>
>
>  --
>  John C. McCabe-Dansted
>  PhD Student
>  University of Western Australia
>

Thanks a lot for your testing efforts.

- Nitin


More information about the linux-mm-cc mailing list