2.5.60-cfq with contest
Here's a quick set of the relevant results with the untuned CFQ by Jens:
ctar_load:
Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 99 78.8 1.0 4.0 1.25 2.5.60-cfq 1 101 78.2 1.0 4.0 1.28 xtar_load: Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 101 76.2 1.0 5.0 1.28 2.5.60-cfq 1 115 66.1 1.0 4.3 1.46 io_load: Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 139 54.7 29.0 12.1 1.76 2.5.60-cfq 1 606 12.7 212.0 21.6 7.67 io_other: Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 90 83.3 10.8 5.5 1.14 2.5.60-cfq 1 89 84.3 10.8 5.6 1.13 read_load: Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 103 74.8 6.2 6.8 1.30 2.5.60-cfq 1 103 76.7 6.9 5.8 1.30 list_load: Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 95 80.0 0.0 6.3 1.20 2.5.60-cfq 1 95 81.1 0.0 6.3 1.20
Write based loads hurt. No breakages, but needs tuning.
Con
Quoted 1 line · Con Kolivas
Here's a quick set of the relevant results with the untuned CFQ by Jens:
Thanks!
Quoted 25 lines
ctar_load:Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 99 78.8 1.0 4.0 1.25 2.5.60-cfq 1 101 78.2 1.0 4.0 1.28 xtar_load: Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 101 76.2 1.0 5.0 1.28 2.5.60-cfq 1 115 66.1 1.0 4.3 1.46 io_load: Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 139 54.7 29.0 12.1 1.76 2.5.60-cfq 1 606 12.7 212.0 21.6 7.67 io_other: Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 90 83.3 10.8 5.5 1.14 2.5.60-cfq 1 89 84.3 10.8 5.6 1.13 read_load: Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 103 74.8 6.2 6.8 1.30 2.5.60-cfq 1 103 76.7 6.9 5.8 1.30 list_load: Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 95 80.0 0.0 6.3 1.20 2.5.60-cfq 1 95 81.1 0.0 6.3 1.20
Quoted 1 line
Write based loads hurt. No breakages, but needs tuning.
That's not even as bad as I had feared. I'll try to do some tuning with contest locally.
Signature · 1 line
Jens Axboe
Quoted 4 lines · Jens Axboe
Quoted 1 line
Write based loads hurt. No breakages, but needs tuning.
That's not even as bad as I had feared. I'll try to do some tuning with contest locally.
Here are my results, for 2.5.60 vanilla, 2.5.60 + cfq with default quantum of 16 (what you tested, too), and 2.5.60 + cfq without quantum setting. The latter should be the fairest, only moves one request from the pending queues.
no_load:
Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 31 177.4 0 0.0 1.00 2.5.60-cfq0 2 31 174.2 0 0.0 1.00 2.5.60-cfq16 2 31 177.4 0 0.0 1.00 cacherun: Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 29 182.8 0 0.0 0.94 2.5.60-cfq0 2 28 192.9 0 0.0 0.90 2.5.60-cfq16 2 29 182.8 0 0.0 0.94 process_load: Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 38 142.1 12 47.4 1.23 2.5.60-cfq0 2 41 129.3 16 61.0 1.32 2.5.60-cfq16 2 37 145.9 12 43.2 1.19 ctar_load: Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 38 147.4 0 0.0 1.23 2.5.60-cfq0 2 36 155.6 0 0.0 1.16 2.5.60-cfq16 2 36 155.6 0 0.0 1.16 xtar_load: Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 40 140.0 0 2.5 1.29 2.5.60-cfq0 2 37 148.6 0 2.7 1.19 2.5.60-cfq16 2 40 137.5 0 2.5 1.29 io_load: Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 93 61.3 2 14.0 3.00 2.5.60-cfq0 4 103 54.4 2 12.6 3.32 2.5.60-cfq16 2 264 21.6 12 19.9 8.52 read_load: Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 40 140.0 0 5.0 1.29 2.5.60-cfq0 2 39 143.6 0 5.1 1.26 2.5.60-cfq16 2 40 140.0 0 5.0 1.29 list_load: Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 35 157.1 0 8.6 1.13 2.5.60-cfq0 2 35 160.0 0 8.6 1.13 2.5.60-cfq16 2 35 160.0 0 14.3 1.13 mem_load: Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 50 116.0 75 10.0 1.61 2.5.60-cfq0 2 57 101.8 78 8.8 1.84 2.5.60-cfq16 2 60 96.7 80 8.2 1.94 dbench_load: Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 36 155.6 12693 27.8 1.16 2.5.60-cfq0 1 35 157.1 12013 28.6 1.13 2.5.60-cfq16 2 37 151.4 14356 32.4 1.19
Signature · 1 line
Jens Axboe
Quoted 10 lines · Jens Axboe
Quoted 4 lines · Jens Axboe
Quoted 1 line
Write based loads hurt. No breakages, but needs tuning.
That's not even as bad as I had feared. I'll try to do some tuning with contest locally.
Here are my results, for 2.5.60 vanilla, 2.5.60 + cfq with default quantum of 16 (what you tested, too), and 2.5.60 + cfq without quantum setting. The latter should be the fairest, only moves one request from the pending queues.
Did runs with quantum values of 2,4,8 as well to see how it looks. Often the dbench runs got screwed, perhaps the signalling changes from 2.5.60 is interfering?
dbench_load.c:72: SYSTEM ERROR: No such process : could not kill pid 4842
Anyways, here are the results:
no_load:
Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 31 177.4 0 0.0 1.00 2.5.60-cfq0 2 31 174.2 0 0.0 1.00 2.5.60-cfq16 2 31 177.4 0 0.0 1.00 2.5.60-cfq4 1 32 171.9 0 0.0 1.00 2.5.60-cfq8 2 31 174.2 0 0.0 1.00 cacherun: Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 29 182.8 0 0.0 0.94 2.5.60-cfq0 2 28 192.9 0 0.0 0.90 2.5.60-cfq16 2 29 182.8 0 0.0 0.94 2.5.60-cfq4 1 29 186.2 0 0.0 0.91 2.5.60-cfq8 2 29 182.8 0 0.0 0.94 process_load: Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 38 142.1 12 47.4 1.23 2.5.60-cfq0 2 41 129.3 16 61.0 1.32 2.5.60-cfq16 2 37 145.9 12 43.2 1.19 2.5.60-cfq4 1 36 150.0 11 44.4 1.12 2.5.60-cfq8 2 38 142.1 13 47.4 1.23 ctar_load: Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 38 147.4 0 0.0 1.23 2.5.60-cfq0 2 36 155.6 0 0.0 1.16 2.5.60-cfq16 2 36 155.6 0 0.0 1.16 2.5.60-cfq4 1 36 155.6 0 0.0 1.12 2.5.60-cfq8 2 37 151.4 0 0.0 1.19 xtar_load: Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 40 140.0 0 2.5 1.29 2.5.60-cfq0 2 37 148.6 0 2.7 1.19 2.5.60-cfq16 2 40 137.5 0 2.5 1.29 2.5.60-cfq4 1 37 148.6 0 2.7 1.16 2.5.60-cfq8 2 38 147.4 0 2.6 1.23 io_load: Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 93 61.3 2 14.0 3.00 2.5.60-cfq0 4 103 54.4 2 12.6 3.32 2.5.60-cfq16 2 264 21.6 12 19.9 8.52 2.5.60-cfq4 1 97 57.7 3 15.5 3.03 2.5.60-cfq8 2 135 42.2 5 16.3 4.35 read_load: Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 40 140.0 0 5.0 1.29 2.5.60-cfq0 2 39 143.6 0 5.1 1.26 2.5.60-cfq16 2 40 140.0 0 5.0 1.29 2.5.60-cfq4 1 39 143.6 0 5.1 1.22 2.5.60-cfq8 2 40 140.0 0 5.0 1.29 list_load: Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 35 157.1 0 8.6 1.13 2.5.60-cfq0 2 35 160.0 0 8.6 1.13 2.5.60-cfq16 2 35 160.0 0 14.3 1.13 2.5.60-cfq4 1 36 155.6 0 8.3 1.12 2.5.60-cfq8 2 35 160.0 0 11.4 1.13 mem_load: Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 50 116.0 75 10.0 1.61 2.5.60-cfq0 2 57 101.8 78 8.8 1.84 2.5.60-cfq16 2 60 96.7 80 8.2 1.94 2.5.60-cfq4 1 52 111.5 76 9.4 1.62 2.5.60-cfq8 2 50 114.0 75 9.8 1.61 dbench_load: Kernel [runs] Time CPU% Loads LCPU% Ratio 2.5.60 2 36 155.6 12693 27.8 1.16 2.5.60-cfq0 1 35 157.1 12013 28.6 1.13 2.5.60-cfq16 2 37 151.4 14356 32.4 1.19 2.5.60-cfq8 1 35 157.1 12174 31.4 1.13
As I initialy expected, without having data to back it up, a non-zero quantum value helps. 16 was too much though, 4 looks a good choice. At least here.
Signature · 1 line
Jens Axboe
--- Jens Axboe <[email protected]> wrote: > On Tue, Feb
Quoted 156 lines · Jens Axboe
Quoted 2 lines · Jens Axboe
Quoted 1 line
Quoted 1 line
Write based loads hurt. No breakages, but
needs tuning.
Quoted 2 lines
Quoted 2 lines
That's not even as bad as I had feared. I'll try
to do some tuning with
Quoted 3 lines
Quoted 1 line
contest locally.
Here are my results, for 2.5.60 vanilla, 2.5.60 +
cfq with default
Quoted 1 line
quantum of 16 (what you tested, too), and 2.5.60 +
cfq without quantum
Quoted 1 line
setting. The latter should be the fairest, only
moves one request from
Quoted 1 line
the pending queues.
Did runs with quantum values of 2,4,8 as well to see how it looks. Often
the dbench runs got screwed, perhaps the signalling changes from 2.5.60
is interfering?dbench_load.c:72: SYSTEM ERROR: No such process :
could not kill pid 4842Anyways, here are the results:
no_load:
Kernel [runs] Time CPU% Loads
LCPU% Ratio
2.5.60 2 31 177.4 0 0.0
1.00
2.5.60-cfq0 2 31 174.2 0 0.0
1.00
2.5.60-cfq16 2 31 177.4 0 0.0
1.00
2.5.60-cfq4 1 32 171.9 0 0.0
1.00
2.5.60-cfq8 2 31 174.2 0 0.0
1.00
cacherun:
Kernel [runs] Time CPU% Loads
LCPU% Ratio
2.5.60 2 29 182.8 0 0.0
0.94
2.5.60-cfq0 2 28 192.9 0 0.0
0.90
2.5.60-cfq16 2 29 182.8 0 0.0
0.94
2.5.60-cfq4 1 29 186.2 0 0.0
0.91
2.5.60-cfq8 2 29 182.8 0 0.0
0.94process_load:
Kernel [runs] Time CPU% Loads
LCPU% Ratio
2.5.60 2 38 142.1 12 47.4
1.23
2.5.60-cfq0 2 41 129.3 16 61.0
1.32
2.5.60-cfq16 2 37 145.9 12 43.2
1.19
2.5.60-cfq4 1 36 150.0 11 44.4
1.12
2.5.60-cfq8 2 38 142.1 13 47.4
1.23ctar_load:
Kernel [runs] Time CPU% Loads
LCPU% Ratio
2.5.60 2 38 147.4 0 0.0
1.23
2.5.60-cfq0 2 36 155.6 0 0.0
1.16
2.5.60-cfq16 2 36 155.6 0 0.0
1.16
2.5.60-cfq4 1 36 155.6 0 0.0
1.12
2.5.60-cfq8 2 37 151.4 0 0.0
1.19xtar_load:
Kernel [runs] Time CPU% Loads
LCPU% Ratio
2.5.60 2 40 140.0 0 2.5
1.29
2.5.60-cfq0 2 37 148.6 0 2.7
1.19
2.5.60-cfq16 2 40 137.5 0 2.5
1.29
2.5.60-cfq4 1 37 148.6 0 2.7
1.16
2.5.60-cfq8 2 38 147.4 0 2.6
1.23io_load:
Kernel [runs] Time CPU% Loads
LCPU% Ratio
2.5.60 2 93 61.3 2 14.0
3.00
2.5.60-cfq0 4 103 54.4 2 12.6
3.32
2.5.60-cfq16 2 264 21.6 12 19.9
8.52
2.5.60-cfq4 1 97 57.7 3 15.5
3.03
2.5.60-cfq8 2 135 42.2 5 16.3
4.35read_load:
Kernel [runs] Time CPU% Loads
LCPU% Ratio
2.5.60 2 40 140.0 0 5.0
1.29
2.5.60-cfq0 2 39 143.6 0 5.1
1.26
2.5.60-cfq16 2 40 140.0 0 5.0
1.29
2.5.60-cfq4 1 39 143.6 0 5.1
1.22
2.5.60-cfq8 2 40 140.0 0 5.0
1.29list_load:
Kernel [runs] Time CPU% Loads
LCPU% Ratio
2.5.60 2 35 157.1 0 8.6
1.13
2.5.60-cfq0 2 35 160.0 0 8.6
1.13
2.5.60-cfq16 2 35 160.0 0 14.3
1.13
2.5.60-cfq4 1 36 155.6 0 8.3
1.12
2.5.60-cfq8 2 35 160.0 0 11.4
1.13mem_load:
Kernel [runs] Time CPU% Loads
LCPU% Ratio
2.5.60 2 50 116.0 75 10.0
1.61
2.5.60-cfq0 2 57 101.8 78 8.8
1.84
2.5.60-cfq16 2 60 96.7 80 8.2
1.94
2.5.60-cfq4 1 52 111.5 76 9.4
1.62
2.5.60-cfq8 2 50 114.0 75 9.8
1.61dbench_load:
Kernel [runs] Time CPU% Loads
LCPU% Ratio
2.5.60 2 36 155.6 12693 27.8
1.16
2.5.60-cfq0 1 35 157.1 12013 28.6
1.13
2.5.60-cfq16 2 37 151.4 14356 32.4
1.19
2.5.60-cfq8 1 35 157.1 12174 31.4
1.13
Quoted 5 lines
As I initialy expected, without having data to back it up, a non-zero
quantum value helps. 16 was too much though, 4 looks a good choice. At
least here.
They look pretty consistent with my results (does any kernel hacker have a uniprocessor machine!?) Impressive that changing a single setting without any other tuning can almost get the same performance as the "tuned up" current scheduler.
You can choose the loads when running contest by just appending the loadnames:
contest no_load io_load xtar_load ctar_load They all need at least one no_load to compare against.
The dbench issue has been fixed by the signal patch from Linus.
Con
http://greetings.yahoo.com.au - Yahoo! Greetings
- Send some online love this Valentine's Day.