Here are the latest contest (http://contest.kolivas.net) results with the osdl
hardware (http://www.osdl.org) in both UP and SMP. This is with a pre- version
of contest so includes cacherun and dbench_load and the load accounting is much
improved.
UP results
noload:
Kernel [runs] Time CPU% Loads LCPU% Ratio
2.4.20 [5] 67.3 97 0 0 1.01
2.5.49 [5] 70.0 96 0 0 1.05
2.5.50 [5] 69.9 96 0 0 1.05
2.5.50-mm1 [5] 71.4 94 0 0 1.07
cacherun:
Kernel [runs] Time CPU% Loads LCPU% Ratio
2.4.20 [5] 65.7 99 0 0 0.98
2.5.49 [5] 67.4 99 0 0 1.01
2.5.50 [5] 67.3 99 0 0 1.01
2.5.50-mm1 [5] 67.8 99 0 0 1.02
process_load:
Kernel [runs] Time CPU% Loads LCPU% Ratio
2.4.20 [5] 108.1 58 84 40 1.62
2.5.49 [5] 85.2 79 17 20 1.28
2.5.50 [5] 84.8 79 17 19 1.27
2.5.50-mm1 [5] 86.6 78 18 20 1.30
dbench_load:
Kernel [runs] Time CPU% Loads LCPU% Ratio
2.4.20 [5] 207.2 32 2 46 3.10
2.5.49 [5] 210.5 37 2 50 3.15
2.5.50 [5] 189.2 40 2 49 2.83
2.5.50-mm1 [5] 243.3 34 2 51 3.64
ctar_load:
Kernel [runs] Time CPU% Loads LCPU% Ratio
2.4.20 [5] 85.4 83 2 9 1.28
2.5.49 [5] 106.1 82 2 9 1.59
2.5.50 [5] 107.5 81 3 9 1.61
2.5.50-mm1 [5] 88.0 83 1 4 1.32
xtar_load:
Kernel [runs] Time CPU% Loads LCPU% Ratio
2.4.20 [5] 107.6 64 2 8 1.61
2.5.49 [5] 184.8 70 3 8 2.77
2.5.50 [5] 189.5 61 4 9 2.84
2.5.50-mm1 [5] 104.9 70 1 6 1.57
io_load:
Kernel [runs] Time CPU% Loads LCPU% Ratio
2.4.20 [5] 203.4 33 40 15 3.05
2.5.49 [5] 127.4 57 14 13 1.91
2.5.50 [5] 142.6 54 19 14 2.14
2.5.50-mm1 [5] 174.2 46 24 15 2.61
io_other:
Kernel [runs] Time CPU% Loads LCPU% Ratio
2.4.20 [5] 120.3 56 24 16 1.80
2.5.49 [5] 97.4 75 7 11 1.46
2.5.50 [5] 106.9 69 10 11 1.60
2.5.50-mm1 [5] 101.8 70 9 11 1.52
read_load:
Kernel [runs] Time CPU% Loads LCPU% Ratio
2.4.20 [5] 88.8 82 16 4 1.33
2.5.49 [5] 88.2 80 15 6 1.32
2.5.50 [5] 88.5 80 15 7 1.33
2.5.50-mm1 [5] 86.6 80 3 2 1.30
list_load:
Kernel [runs] Time CPU% Loads LCPU% Ratio
2.4.20 [5] 75.7 88 0 8 1.13
2.5.49 [5] 81.4 85 0 8 1.22
2.5.50 [5] 81.2 85 0 8 1.22
2.5.50-mm1 [5] 82.4 84 0 7 1.23
mem_load:
Kernel [runs] Time CPU% Loads LCPU% Ratio
2.4.20 [5] 84.8 80 44 2 1.27
2.5.49 [5] 98.1 76 43 2 1.47
2.5.50 [5] 98.3 76 44 2 1.47
2.5.50-mm1 [5] 116.9 67 47 1 1.75
SMP Results
noload:
Kernel [runs] Time CPU% Loads LCPU% Ratio
2.4.20 [5] 36.9 189 0 0 1.02
2.5.49 [6] 39.3 181 0 0 1.09
2.5.50 [5] 39.3 180 0 0 1.09
2.5.50-mm1 [6] 39.4 181 0 0 1.09
cacherun:
Kernel [runs] Time CPU% Loads LCPU% Ratio
2.4.20 [5] 35.8 194 0 0 0.99
2.5.49 [6] 36.6 194 0 0 1.01
2.5.50 [5] 36.5 194 0 0 1.01
2.5.50-mm1 [6] 36.6 194 0 0 1.01
process_load:
Kernel [runs] Time CPU% Loads LCPU% Ratio
2.4.20 [5] 51.0 138 20 59 1.41
2.5.49 [6] 50.0 141 11 52 1.38
2.5.50 [5] 47.8 148 10 46 1.32
2.5.50-mm1 [5] 47.6 150 8 43 1.31
dbench_load:
Kernel [runs] Time CPU% Loads LCPU% Ratio
2.4.20 [5] 144.8 49 0 38 4.00
2.5.49 [5] 119.8 96 0 26 3.31
2.5.50 [5] 199.8 101 0 24 5.52
2.5.50-mm1 [5] 164.3 67 0 29 4.54
ctar_load:
Kernel [runs] Time CPU% Loads LCPU% Ratio
2.4.20 [5] 43.1 169 1 11 1.19
2.5.49 [5] 53.8 161 1 10 1.49
2.5.50 [5] 54.6 157 1 10 1.51
2.5.50-mm1 [5] 51.3 155 0 4 1.42
xtar_load:
Kernel [runs] Time CPU% Loads LCPU% Ratio
2.4.20 [5] 63.5 114 1 12 1.75
2.5.49 [5] 72.9 132 1 10 2.01
2.5.50 [5] 116.2 103 2 10 3.21
2.5.50-mm1 [5] 83.9 111 1 9 2.32
io_load:
Kernel [runs] Time CPU% Loads LCPU% Ratio
2.4.20 [5] 164.9 45 31 21 4.55
2.5.49 [5] 75.5 110 9 18 2.09
2.5.50 [5] 87.6 102 14 22 2.42
2.5.50-mm1 [5] 99.0 92 14 21 2.73
io_other:
Kernel [runs] Time CPU% Loads LCPU% Ratio
2.4.20 [5] 89.6 86 17 21 2.47
2.5.49 [5] 64.2 130 8 19 1.77
2.5.50 [5] 59.3 139 7 18 1.64
2.5.50-mm1 [5] 70.5 125 10 22 1.95
read_load:
Kernel [runs] Time CPU% Loads LCPU% Ratio
2.4.20 [5] 51.0 143 7 7 1.41
2.5.49 [5] 49.1 152 5 7 1.36
2.5.50 [5] 49.3 151 5 7 1.36
2.5.50-mm1 [5] 52.1 142 2 3 1.44
list_load:
Kernel [runs] Time CPU% Loads LCPU% Ratio
2.4.20 [5] 40.4 175 0 8 1.12
2.5.49 [5] 43.4 167 0 8 1.20
2.5.50 [5] 43.4 167 0 8 1.20
2.5.50-mm1 [5] 44.0 167 0 7 1.22
mem_load:
Kernel [runs] Time CPU% Loads LCPU% Ratio
2.4.20 [5] 52.0 139 37 3 1.44
2.5.49 [5] 62.5 145 35 3 1.73
2.5.50 [5] 63.3 141 36 3 1.75
2.5.50-mm1 [5] 67.1 126 39 3 1.85
Further information will be available at
http://www.osdl.org/projects/ctdevel/results/ (short lag time for the web server
to catch up).
Note that some of these runs have pathological cases so looking at the full load
information (eg io_load.log) will show you these. Thats why average of 5 runs is
performed. I may move to 7 runs in the future.
Con
On 03-Dec-2002 Con Kolivas wrote:
>
> UP results
>
> process_load:
> Kernel [runs] Time CPU% Loads LCPU% Ratio
> 2.4.20 [5] 108.1 58 84 40 1.62
> 2.5.50-mm1 [5] 86.6 78 18 20 1.30
Hm, load task gets half cpu time, but it goes 5 times slower
in 2.5.x. Why ? You can see a similar behaviour in other
tests too.
Bye.
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1
On Wed, 4 Dec 2002 07:43 pm, Giuliano Pochini wrote:
> On 03-Dec-2002 Con Kolivas wrote:
> > UP results
> >
> > process_load:
> > Kernel [runs] Time CPU% Loads LCPU% Ratio
> > 2.4.20 [5] 108.1 58 84 40 1.62
> > 2.5.50-mm1 [5] 86.6 78 18 20 1.30
>
> Hm, load task gets half cpu time, but it goes 5 times slower
> in 2.5.x. Why ? You can see a similar behaviour in other
> tests too.
You have to dig deeper to understand why. The time taken to compile the kernel
takes a fixed amount of cpu time. In the presence of a load, it takes longer
in wall clock time, but still takes about the same amount of cpu time. The
amount of extra wall clock time will basically be used for the load,
scheduling, io etc. Now the absolute extra wall clock time it takes to
compile the kernel in this load is time it can also be doing the load.
Therefore, if I spend 20 seconds longer to compile the kernel while the load
is running, the load must also get 20 seconds where it can be doing it's
work. Assuming it does 1 load per second, it will do 20 loads. If it takes 40
seconds longer to compile the kernel, the load gets 40 seconds longer; hence
it can do 40 loads.
Look at the example above and you'll see those numbers. It takes 20 seconds
longer in 2.5.50-mm1 compared to noload, and load gets to do 20 workloads.
2.4.20 takes 40 seconds longer and gets to do 40 workloads. If you take into
account the work done / time they are doing workloads at about the same rate.
Now if one had taken 20 seconds longer than the other and done only the same
amount of work it would be showing serious inefficiencies over and above the
fair scheduling issues which contest is really trying to measure.
Hmm. Not sure if I made that clear enough, but hopefully I got my point
across.
Con
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.0 (GNU/Linux)
iD8DBQE97crOF6dfvkL3i1gRAhdzAKCX8vlHqLQUm+MnzsGAjzP7UPJB4ACbB4um
XyNURkBWQwIC7xAvgkTwmpY=
=4mwU
-----END PGP SIGNATURE-----