scheduler tunables with contest - interactive_delta

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

osdl hardware, contest benchmark, 2.5.52-mm2 varying the interactive delta:

I've cut it down to the two loads that seem to be affected significantly. The other changes are subtle. id1 is interactive_delta = 1 and so on:

io_load:

Kernel [runs]           Time    CPU%    Loads   LCPU%   Ratio
id1 [5]                 78.2    108     9       19      2.16
id2 [5]                 87.9    96      12      19      2.43
id3 [5]                 79.4    105     10      18      2.19
id4 [5]                 84.9    109     12      22      2.34
id5 [7]                 99.8    94      18      24      2.76
id6 [5]                 103.3   104     18      25      2.85
id7 [5]                 104.1   89      17      22      2.87

mem_load:

Kernel [runs]           Time    CPU%    Loads   LCPU%   Ratio
id1 [3]                 78.0    100     34      2       2.15
id2 [5]                 94.3    84      35      2       2.60
id3 [5]                 92.6    85      32      2       2.56
id4 [5]                 65.8    134     39      3       1.82
id5 [5]                 63.0    143     38      3       1.74
id6 [5]                 69.7    129     38      2       1.92
id7 [5]                 90.6    87      32      2       2.50

Seems like io_load likes lower interactive deltas (lower the better?) and mem_load likes high interactive_deltas (sweet spot 5).

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

iD8DBQE+A/CEF6dfvkL3i1gRAoM2AJ45DsfpltAWXNoaXIWmArMRdz2PIgCffpWP A9gVU7M6NBIoGaFYQyx17wE=
=Mayt
-----END PGP SIGNATURE-----

On Saturday 21 December 2002 05:39, Con Kolivas wrote:

Hi Con,

Yes, seems so. I think this is a good thing for autoregulating 8-)

ciao, Marc

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

Exactly my thoughts. Increasingly as the numbers are rolling out it is clear the defaults seem to be pretty darn good (thanks mingo), but for -ck at least I wont be able to resist and will autoregulate them once I have a full set of numbers to play with. I'll get subsets of numbers (too messy to post them all here) and decide what to use for my ranges. It's my impression so far that the desktop experience with static numbers can only be improved upon by dropping the max timeslice to resemble that of the old scheduler. Perhaps a max timeslice of around 150ms.

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

iD8DBQE+BGFVF6dfvkL3i1gRAkQBAJsH5MLDRvSpa2VIY+u4Up2FZhdkUQCgg8sq FAQx+63jqkrR1IUHIA3zZVQ=
=jWrY
-----END PGP SIGNATURE-----