[BENCHMARK] scheduler tunables with contest - exit_weight

[BENCHMARK] scheduler tunables with contest - exit_weight

[Con Kolivas]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

osdl hardware, contest results, 2.5.52-mm2 with scheduler tunable - exit 
weight (ew1= exit weight ==1 and so on)

io_load:
Kernel [runs]           Time    CPU%    Loads   LCPU%   Ratio
ew0 [5]                 105.3   90      16      22      2.91
ew1 [5]                 86.4    97      12      18      2.39
ew2 [5]                 74.9    109     9       18      2.07
ew3 [5]                 84.2    100     11      19      2.33
ew4 [5]                 83.8    102     10      18      2.31
ew5 [5]                 89.9    93      12      20      2.48
ew6 [5]                 97.5    88      13      20      2.69
ew7 [5]                 89.2    95      12      20      2.46

mem_load:
Kernel [runs]           Time    CPU%    Loads   LCPU%   Ratio
ew0 [5]                 89.2    90      32      2       2.46
ew1 [5]                 90.5    88      33      2       2.50
ew2 [5]                 87.2    91      32      2       2.41
ew3 [5]                 88.7    92      32      2       2.45
ew4 [5]                 91.5    88      32      2       2.53
ew5 [5]                 87.7    89      32      2       2.42
ew6 [5]                 90.2    87      32      2       2.49
ew7 [5]                 88.6    91      32      2       2.45

It seems with these results at least, benefit is gained from it being on 
versus off. The actual value seems not important wrt contest.

Con
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.0 (GNU/Linux)

iD8DBQE+BNqcF6dfvkL3i1gRAqeeAJ9b3OLhmb0737HfbJG1N9QMOou8gQCeJ6OJ
p2IhcdJ3QeBx9k3QX5+6yzk=
=Iaf8
-----END PGP SIGNATURE-----

Re: [BENCHMARK] scheduler tunables with contest - exit_weight

[Marc-Christian Petersen]

On Saturday 21 December 2002 22:18, Con Kolivas wrote:

Hey Con,

> osdl hardware, contest results, 2.5.52-mm2 with scheduler tunable - exit
> weight (ew1= exit weight ==1 and so on)
Can you please try another thing?

kernel/sched.c

        /*
         * If the child was a (relative-) CPU hog then decrease
         * the sleep_avg of the parent as well.  
         */
        if (p->sleep_avg < p->parent->sleep_avg)
                p->parent->sleep_avg = (p->parent->sleep_avg * exit_weight +
                        p->sleep_avg) / (exit_weight + 1);

Remove these please and run again.

ciao, Marc

Re: [BENCHMARK] scheduler tunables with contest - exit_weight

[Con Kolivas]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

On Mon, 23 Dec 2002 11:49 am, Marc-Christian Petersen wrote:
> On Saturday 21 December 2002 22:18, Con Kolivas wrote:
>
> Hey Con,
>
> > osdl hardware, contest results, 2.5.52-mm2 with scheduler tunable - exit
> > weight (ew1= exit weight ==1 and so on)
>
> Can you please try another thing?
>
> kernel/sched.c
>
>         /*
>          * If the child was a (relative-) CPU hog then decrease
>          * the sleep_avg of the parent as well.
>          */
>         if (p->sleep_avg < p->parent->sleep_avg)
>                 p->parent->sleep_avg = (p->parent->sleep_avg * exit_weight
> + p->sleep_avg) / (exit_weight + 1);
>
> Remove these please and run again.

Sure why not

noload:
Kernel [runs]           Time    CPU%    Loads   LCPU%   Ratio
2.5.52-mm2 [7]          39.2    181     0       0       1.08
2552mm2noew [5]         39.6    180     0       0       1.09

cacherun:
Kernel [runs]           Time    CPU%    Loads   LCPU%   Ratio
2.5.52-mm2 [7]          36.5    194     0       0       1.01
2552mm2noew [5]         36.5    194     0       0       1.01

process_load:
Kernel [runs]           Time    CPU%    Loads   LCPU%   Ratio
2.5.52-mm2 [7]          46.5    152     8       41      1.28
2552mm2noew [5]         49.1    143     10      50      1.36

ctar_load:
Kernel [runs]           Time    CPU%    Loads   LCPU%   Ratio
2.5.52-mm2 [7]          52.8    154     1       10      1.46
2552mm2noew [5]         57.7    160     1       10      1.59

xtar_load:
Kernel [runs]           Time    CPU%    Loads   LCPU%   Ratio
2.5.52-mm2 [7]          76.1    124     1       8       2.10
2552mm2noew [5]         70.1    128     1       8       1.94

io_load:
Kernel [runs]           Time    CPU%    Loads   LCPU%   Ratio
2.5.52-mm2 [7]          74.5    112     11      20      2.06
2552mm2noew [5]         67.3    117     8       17      1.86

io_other:
Kernel [runs]           Time    CPU%    Loads   LCPU%   Ratio
2.5.52-mm2 [7]          59.9    134     6       18      1.65
2552mm2noew [5]         66.2    131     9       21      1.83

read_load:
Kernel [runs]           Time    CPU%    Loads   LCPU%   Ratio
2.5.52-mm2 [7]          50.5    147     5       6       1.39
2552mm2noew [5]         50.2    149     5       6       1.39

list_load:
Kernel [runs]           Time    CPU%    Loads   LCPU%   Ratio
2.5.52-mm2 [7]          43.7    167     0       9       1.21
2552mm2noew [5]         43.5    166     0       9       1.20

mem_load:
Kernel [runs]           Time    CPU%    Loads   LCPU%   Ratio
2.5.52-mm2 [7]          66.0    141     39      3       1.82
2552mm2noew [5]         64.7    136     38      3       1.79

Con
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.0 (GNU/Linux)

iD8DBQE+B7yAF6dfvkL3i1gRAhjaAJ4r0rEcvy95o28bw7WoSBoe7ZfmAgCgpHJS
HFJaBj8sBtA1MyzK/7qptQg=
=N3fU
-----END PGP SIGNATURE-----

Re: [BENCHMARK] scheduler tunables with contest - exit_weight

[Con Kolivas]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

On Thu, 26 Dec 2002 12:24 am, Denis Vlasenko wrote:
> On 21 December 2002 19:18, Con Kolivas wrote:
> > -----BEGIN PGP SIGNED MESSAGE-----
> > Hash: SHA1
> >
> > osdl hardware, contest results, 2.5.52-mm2 with scheduler tunable -
> > exit weight (ew1= exit weight ==1 and so on)
> >
> > io_load:
> > Kernel [runs]           Time    CPU%    Loads   LCPU%   Ratio
> > ew0 [5]                 105.3   90      16      22      2.91
> > ew1 [5]                 86.4    97      12      18      2.39
> > ew2 [5]                 74.9    109     9       18      2.07
> > ew3 [5]                 84.2    100     11      19      2.33
> > ew4 [5]                 83.8    102     10      18      2.31
> > ew5 [5]                 89.9    93      12      20      2.48
> > ew6 [5]                 97.5    88      13      20      2.69
> > ew7 [5]                 89.2    95      12      20      2.46
>
> In spite of worrying reports of decreasing single task performance,
> does it make sense to add "null_load" test? ;)

I've simplified the data. There is no significant difference in the no_load 
groups with changes to the scheduler tunables over useful ranges. 

Con
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.0 (GNU/Linux)

iD8DBQE+CXyRF6dfvkL3i1gRAtSyAJ9XThpp5iCI1FcjDxVOESbm5ialywCgg7Vb
HN+jWurjIwXngqCUOmDWhh0=
=7Iy8
-----END PGP SIGNATURE-----

Re: [BENCHMARK] scheduler tunables with contest - exit_weight

[Denis Vlasenko]

On 21 December 2002 19:18, Con Kolivas wrote:
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
>
> osdl hardware, contest results, 2.5.52-mm2 with scheduler tunable -
> exit weight (ew1= exit weight ==1 and so on)
>
> io_load:
> Kernel [runs]           Time    CPU%    Loads   LCPU%   Ratio
> ew0 [5]                 105.3   90      16      22      2.91
> ew1 [5]                 86.4    97      12      18      2.39
> ew2 [5]                 74.9    109     9       18      2.07
> ew3 [5]                 84.2    100     11      19      2.33
> ew4 [5]                 83.8    102     10      18      2.31
> ew5 [5]                 89.9    93      12      20      2.48
> ew6 [5]                 97.5    88      13      20      2.69
> ew7 [5]                 89.2    95      12      20      2.46

In spite of worrying reports of decreasing single task performance,
does it make sense to add "null_load" test? ;)
--
vda