Here is a family of contest benchmarks using the osdl hardware in uniprocessor
mode on 2.5.53-mm1 while varying vm swappiness. s020 is vm swappiness=20 and
so on:
noload:
Kernel [runs] Time CPU% Loads LCPU% Ratio
s000 [3] 71.1 95 0 0 1.07
s020 [5] 71.9 95 0 0 1.08
s040 [5] 71.7 95 0 0 1.07
s060 [5] 71.3 96 0 0 1.07
s080 [5] 71.3 95 0 0 1.07
s100 [5] 71.6 95 0 0 1.07
cacherun:
Kernel [runs] Time CPU% Loads LCPU% Ratio
s000 [3] 68.4 99 0 0 1.02
s020 [5] 68.8 99 0 0 1.03
s040 [5] 68.7 99 0 0 1.03
s060 [5] 68.6 99 0 0 1.03
s080 [5] 68.5 99 0 0 1.03
s100 [5] 68.7 99 0 0 1.03
process_load:
Kernel [runs] Time CPU% Loads LCPU% Ratio
s000 [3] 119.0 57 49 41 1.78
s020 [5] 119.4 57 49 41 1.79
s040 [5] 118.6 57 48 41 1.78
s060 [5] 117.6 57 47 41 1.76
s080 [5] 119.5 57 49 41 1.79
s100 [5] 118.4 58 48 40 1.77
dbench_load:
Kernel [runs] Time CPU% Loads LCPU% Ratio
s000 [3] 191.6 41 1 43 2.87
s020 [5] 195.5 40 1 44 2.93
s040 [5] 197.9 41 1 43 2.96
s060 [5] 331.4 32 0 23 4.96
s080 [5] 439.4 24 0 10 6.58
s100 [5] 883.6 13 1 9 13.24
The first of the massive effect of changing this value. A recurring theme is
large file writes with a swappy kernel seems to waste time swapping data
because of the IO data. Obviously none of these IO loads use the o_direct
option.
ctar_load:
Kernel [runs] Time CPU% Loads LCPU% Ratio
s000 [3] 113.4 81 3 10 1.70
s020 [5] 103.3 80 2 9 1.55
s040 [5] 110.0 79 3 9 1.65
s060 [5] 103.0 80 3 9 1.54
s080 [5] 104.4 80 2 8 1.56
s100 [5] 100.2 80 2 8 1.50
Slightly slower with an unswappy kernel
xtar_load:
Kernel [runs] Time CPU% Loads LCPU% Ratio
s000 [3] 139.2 65 3 8 2.09
s020 [5] 128.7 70 2 7 1.93
s040 [5] 152.8 59 3 7 2.29
s060 [5] 137.0 64 2 6 2.05
s080 [5] 124.6 66 2 6 1.87
s100 [5] 127.6 66 2 6 1.91
Up and down, no firm pattern
io_load:
Kernel [runs] Time CPU% Loads LCPU% Ratio
s000 [3] 121.8 63 21 17 1.82
s020 [5] 125.9 61 22 17 1.89
s040 [5] 130.1 59 22 17 1.95
s080 [5] 174.6 47 27 15 2.62
s100 [5] 208.1 42 25 11 3.12
Increase the swappiness, increase the kernel compile time as dbench_load does
io_other:
Kernel [runs] Time CPU% Loads LCPU% Ratio
s000 [3] 103.6 73 18 17 1.55
s020 [5] 135.4 62 25 18 2.03
s040 [5] 157.7 57 28 18 2.36
s060 [5] 188.1 48 32 16 2.82
s080 [5] 246.2 37 38 15 3.69
s100 [5] 378.8 24 45 11 5.67
Another massive change.
read_load:
Kernel [runs] Time CPU% Loads LCPU% Ratio
s000 [3] 89.7 80 15 7 1.34
s020 [5] 91.1 79 13 6 1.36
s040 [5] 90.6 79 12 6 1.36
s060 [5] 90.3 79 12 6 1.35
s080 [5] 90.3 79 12 6 1.35
s100 [5] 92.2 78 10 5 1.38
list_load:
Kernel [runs] Time CPU% Loads LCPU% Ratio
s000 [3] 81.8 85 0 9 1.23
s020 [5] 81.9 85 0 9 1.23
s040 [5] 82.0 85 0 9 1.23
s060 [5] 82.2 85 0 8 1.23
s080 [5] 82.7 85 0 8 1.24
s100 [5] 83.1 85 0 8 1.24
mem_load:
Kernel [runs] Time CPU% Loads LCPU% Ratio
s000 [3] 233.2 32 65 1 3.49
s020 [5] 172.3 42 50 1 2.58
s040 [5] 163.8 46 44 1 2.45
s060 [5] 192.2 39 38 1 2.88
s080 [5] 186.4 42 37 0 2.79
s100 [5] 128.2 57 37 1 1.92
Here it seems only the extreme values make a difference. Un unswappy kernel
makes it slow down significantly, whereas a swappy kernel speeds it up
significantly.
Paolo if I was to choose a number from these values I'd suggest lower than 60
rather than higher, BUT that is because the io load effects become a real
problem when the kernel is swappy - don't _really_ know what this means for
the rest of the time. Maybe in the 40-50 range. There seems to be a knee
(bend) in the curve (most noticable in dbench_load) rather than the curves
being linear. That knee I believe simply shows the way the algorithm for
swappiness basically works. I might throw a 50 at the machine as well to see
what that does.
Con
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.0 (GNU/Linux)
iD8DBQE+C+kYF6dfvkL3i1gRAt2HAJ9UJFKO2i5K6LW/7WU+cJ1TBOtlEQCeISK7
7ecGaUImIiJvC1Bz649vmH8=
=UI1t
-----END PGP SIGNATURE-----
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/