On Mon, 2011-02-14 at 11:34 +0000, Ben Hutchings wrote:
> On Mon, 2011-02-14 at 11:23 +0100, Bastian Blank wrote:
> > Hi folks
> >
> > I'd like to drop the i686 non-pae kernel. Currently we have sometimes
> > -686 with PAE; only the normal kernel is without PAE. I'd like to get
> > rid of this problem. Also this enables the use of the NX bit if supported
> > by the CPU.
> >
> > There are some i686 processors without PAE support. This are some of the
> > Pentium M (all of the Banias line and some of the Dothan line) and the
> > Via C3 Nehemiah. All of them are released 2005 and earlier.
>
> Also Geode LX.
>
> Are there any changes we could/should make to the 486 flavour that would
> make it perform better on 686-class processors? Should we consider also
> dropping 486 support and making it a 586 flavour with corresponding
> optimisations?
[...]
Before we make a change, I wanted to do a little testing to see whether
running the 486 flavour is likely to hurt performance on the 686-class
processors without PAE. And I do mean a little testing - I don't want
to spend time constructing complex benchmarks.
On my home server, which has one of the affected processors,
specifically a VIA C3 "Nehemiah", I ran a simple benchmark for
filesystem access:
for i in $(seq 0 100); do \time find /usr >/dev/null; done
and took the 'system' times from all but the first iteration (i.e. only
the cache-hot case).
I then compared these data sets with 'ministat':
x 486
+ 686
+------------------------------------------------------------------------------+
| x + |
| x + + |
| x + + |
| x + + + |
| x + * + |
| x x x + * + |
| x x x + * + |
| x * x x + * + + |
| x * x * + * + + |
| x * * * + * + + |
| x * * * * * + + + |
| x * * * * * + * + |
| x * * * * * + * + |
| x * * * * * * * + + |
| x x * * * * * * * + + |
| * * * * * * * * * + + |
|x * * * * * * * * * + * +|
| |___________AM___________| |
| |______________A_M____________| |
+------------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 100 0.27 0.38 0.32 0.3192 0.021304905
+ 100 0.28 0.4 0.34 0.336 0.02502524
Difference at 95.0% confidence
0.0168 +/- 0.0064417
5.26316% +/- 2.01808%
(Student's t, pooled s = 0.0232396)
For this limited test, the 486 kernel actually seems to be slightly
faster. Note that this was *not* run on an idle system, so other
activity could affect the measurements a little.
The Pentium M processors are likely to have different performance
characteristics so I would like to see someone test them as well.
It might be worth doing some kind of networking benchmark too.
Ben.
--
Ben Hutchings
Once a job is fouled up, anything done to improve it makes it worse.
Attachment:
signature.asc
Description: This is a digitally signed message part