Skip to contentSkip to conversation

massive_intr.c -- cfs/vanilla/sd-0.40

TEST RESULT [email protected] 5 participants 8 messages Apr 14, 2007 – Apr 16, 2007

lore
Share & tools
lore
#0 lore raw

Hi Ingo,
Did a test with massive_intr.c on a standard linux desktop.
for vanilla, con's Sd-0.40 and cfs.

Test bed is a p4 2.26GHz, 512mb ram
No load on the machine expect for the X server.

Results are stated below

Vanilla Kernel
--------------------
2.6.21-rc6-vanilla #8 SMP Sat Apr 14

[surya@bluegenie tests]$ ./massive_intr 10 10

002469  00000548
002463  00000499
002464  00000026
002465  00000028
002472  00000021
002467  00000018
002470  00000015
002466  00000017
002468  00000024
002471  00000013

top output
top - 13:49:53 up 4 min, 3 users, load average: 1.24, 0.71, 0.30

Tasks:  80 total,  10 running,  70 sleeping,   0 stopped,   0 zombie
Cpu(s): 13.3%us,  1.8%sy,  0.0%ni, 72.1%id, 12.7%wa,  0.1%hi,  0.0%si,
0.0%st
Mem:    506112k total,   440980k used,    65132k free,    31444k buffers
Swap:   522104k total,        0k used,   522104k free,   309560k cached

PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+
COMMAND
2469 surya 16 0 1624 204 132 S 33.8 0.0 0:00.30
massive_intr
2467 surya 22 0 1624 204 132 R 19.9 0.0 0:00.20
massive_intr
2466 surya 23 0 1624 204 132 R 15.9 0.0 0:00.20
massive_intr
2470 surya 24 0 1624 204 132 R 13.9 0.0 0:00.18
massive_intr
2468 surya 19 0 1624 204 132 R 8.0 0.0 0:00.19
massive_intr
2463 surya 19 0 1624 204 132 R 6.0 0.0 0:00.26
massive_intr

    1 root      15   0  2032  640  548 S  0.0  0.1   0:00.72 

bash
<snip>
2462 surya 25 0 1624 428 356 S 0.0 0.1 0:01.63
massive_intr
2464 surya 20 0 1624 204 132 R 0.0 0.0 0:00.21
massive_intr
2465 surya 21 0 1624 204 132 R 0.0 0.0 0:00.20
massive_intr
2471 surya 20 0 1624 204 132 R 0.0 0.0 0:00.10
massive_intr
2472 surya 21 0 1624 204 132 R 0.0 0.0 0:00.20
massive_intr
2473 surya 15 0 2160 900 724 R 0.0 0.2 0:00.00 top

SD-0.40
---------------

2.6.21-rc6-rsdl0.40 #6 SMP Sat Apr 14 13:08:31 IST 2007 i686 i686 i386 GNU/Linux
[root@bluegenie tests]# ./massive_intr 10 10

002504  00000125
002502  00000124
002500  00000123
002499  00000127
002501  00000125
002505  00000127
002506  00000127
002503  00000128
002498  00000124
002507  00000127

top - 13:42:50 up 5 min, 3 users, load average: 2.03, 1.31, 0.61

Tasks:  80 total,  12 running,  68 sleeping,   0 stopped,   0 zombie
Cpu(s): 17.8%us,  1.4%sy,  0.0%ni, 69.8%id, 11.0%wa,  0.1%hi,  0.0%si,
0.0%st
Mem:    506112k total,   441664k used,    64448k free,    31592k buffers
Swap:   522104k total,        0k used,   522104k free,   310152k cached

PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+
COMMAND
2500 root 29 0 1620 204 132 R 11.2 0.0 0:00.28
massive_intr
2501 root 29 0 1620 204 132 R 11.2 0.0 0:00.30
massive_intr
2502 root 29 0 1620 204 132 R 11.2 0.0 0:00.30
massive_intr
2504 root 29 0 1620 204 132 R 11.2 0.0 0:00.30
massive_intr
2498 root 29 0 1620 204 132 R 9.3 0.0 0:00.29
massive_intr
2499 root 29 0 1620 204 132 R 9.3 0.0 0:00.28
massive_intr
2503 root 27 0 1620 204 132 R 9.3 0.0 0:00.27
massive_intr
2505 root 27 0 1620 204 132 R 9.3 0.0 0:00.29
massive_intr
2506 root 27 0 1620 204 132 R 9.3 0.0 0:00.28
massive_intr
2507 root 27 0 1620 204 132 R 9.3 0.0 0:00.28
massive_intr

    1 root       5   0  2036  640  548 S  0.0  0.1   0:00.72

init
<snip>
2497 root 1 0 1620 428 356 S 0.0 0.1 0:01.69
massive_intr
2508 surya 27 0 2164 904 724 R 0.0 0.2 0:00.00 top

CFS-kernel
----------------

2.6.21-rc6-cfs #7 SMP Sat Apr 14

pass1
-----
[surya@bluegenie tests]$ ./massive_intr 10 10

002435  00000120
002439  00000120
002441  00000120
002434  00000120
002436  00000120
002440  00000120
002432  00000120
002437  00000120
002433  00000120
002438  00000120

Felt it is too much fair, will try another pass ;)

[surya@bluegenie tests]$ ./massive_intr 10 10

002961  00000121
002965  00000120
002964  00000121
002959  00000120
002956  00000121
002963  00000121
002960  00000121
002962  00000121
002958  00000122
002957  00000122

top_output
---
top - 13:55:39 up 3 min, 3 users, load average: 0.24, 0.46, 0.22

Tasks:  80 total,  11 running,  69 sleeping,   0 stopped,   0 zombie
Cpu(s):  6.7%us,  2.2%sy,  0.0%ni, 75.3%id, 15.7%wa,  0.1%hi,  0.0%si,
0.0%st
Mem:    506108k total,   438100k used,    68008k free,    31144k buffers
Swap:   522104k total,        0k used,   522104k free,   306932k cached

PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+
COMMAND
2439 surya 20 0 1620 204 132 R 12.7 0.0 0:00.24
massive_intr
2432 surya 20 0 1620 204 132 R 10.9 0.0 0:00.24
massive_intr
2441 surya 20 0 1620 204 132 R 10.9 0.0 0:00.25
massive_intr
2433 surya 20 0 1620 204 132 R 9.1 0.0 0:00.20
massive_intr
2435 surya 20 0 1620 204 132 R 9.1 0.0 0:00.23
massive_intr
2434 surya 20 0 1620 204 132 R 7.2 0.0 0:00.23
massive_intr
2436 surya 20 0 1620 204 132 R 7.2 0.0 0:00.20
massive_intr
2437 surya 20 0 1620 204 132 R 7.2 0.0 0:00.20
massive_intr
2438 surya 20 0 1620 204 132 R 7.2 0.0 0:00.22
massive_intr
2440 surya 20 0 1620 204 132 R 7.2 0.0 0:00.21
massive_intr

    1 root      20   0  2036  640  548 S  0.0  0.1   0:00.72
init               
    2 root      RT   0     0    0    0 S  0.0  0.0   0:00.00 

<snip>
2431 surya 20 0 1620 428 356 S 0.0 0.1 0:01.65
massive_intr
2442 surya 20 0 2160 900 724 R 0.0 0.2 0:00.00
top

Noticed considerable amount of load while running con's kernel. cpu time is very very close to being fair in cfs. Will be trying out ringtest in the next round.

thanks..

Signature · 2 lines
surya.
14/04/2007
#1 lore raw

Hi Surya,

On Sat, Apr 14, 2007 at 02:19:06PM +0530, [email protected] wrote:
(...)
thanks for the interesting results.

Quoted 2 lines

cpu time is very very close to being fair in cfs. Will be trying out ringtest in the next round.

It looks like Ingo can do very good things, he just needs very strong competition to decide to go to the keyboard ;-)

Sorta like lazy sportives who decide to make an effort at the last moment not to lose their first place...

Cheers,
Willy

#2 lore raw
In reply to [email protected]
Quoted 3 lines · * [email protected] <[email protected]> wrote:

Hi Ingo,
Did a test with massive_intr.c on a standard linux desktop.
for vanilla, con's Sd-0.40 and cfs.

thanks!

Quoted 12 lines

[surya@bluegenie tests]$ ./massive_intr 10 10

002435  00000120
002439  00000120
002441  00000120
002434  00000120
002436  00000120
002440  00000120
002432  00000120
002437  00000120
002433  00000120
002438  00000120
Quoted 1 line

Felt it is too much fair, will try another pass ;)

hehe :)

Quoted 11 lines

[surya@bluegenie tests]$ ./massive_intr 10 10

002961  00000121
002965  00000120
002964  00000121
002959  00000120
002956  00000121
002963  00000121
002960  00000121
002962  00000121
002958  00000122
002957  00000122

btw., other schedulers might work better with some more test-time: i'd suggest to use 60 seconds (./massive_intr 10 60) [or maybe more, using more threads] to see long-term fairness effects.

Quoted 1 line

[...] Will be trying out ringtest in the next round.

cool. ringtest.c is intended to be used the following way: start it, it will generate a 99% busy system (but it is using a ring of 100 tasks, where each tasks runs for 100 msecs then sleeps for 1 msec, so every task gets a turn every 10 seconds). If you add a pure CPU hog to the system, for example an infinite shell loop:

while :; do :; done &

then a 'fair' scheduler would give roughly 50% of CPU time to the CPU hog (and the ringtest.c tasks take up the other 50%).

Ingo

#3 lore raw
Quoted 8 lines · Ingo Molnar

cool. ringtest.c is intended to be used the following way: start it, it will generate a 99% busy system (but it is using a ring of 100 tasks, where each tasks runs for 100 msecs then sleeps for 1 msec, so every task gets a turn every 10 seconds). If you add a pure CPU hog to the system, for example an infinite shell loop:
while :; do :; done &
then a 'fair' scheduler would give roughly 50% of CPU time to the CPU hog (and the ringtest.c tasks take up the other 50%).

I've queued up modifying ringtest.c to automatically spawn a CPU hog and then report aggregate CPU bandwidth of the ring and the bandwidth of the CPU hog as work to do at some point. I've no guarantee I'll get to it in a timely fashion, though.

-- wli

#4 lore raw
In reply to Ingo Molnar

At Sat, 14 Apr 2007 14:02:20 +0200,

Quoted 38 lines · Ingo Molnar
Quoted 3 lines · * [email protected] <[email protected]> wrote:

Hi Ingo,
Did a test with massive_intr.c on a standard linux desktop.
for vanilla, con's Sd-0.40 and cfs.

thanks!

Quoted 12 lines

[surya@bluegenie tests]$ ./massive_intr 10 10

002435  00000120
002439  00000120
002441  00000120
002434  00000120
002436  00000120
002440  00000120
002432  00000120
002437  00000120
002433  00000120
002438  00000120
Quoted 1 line

Felt it is too much fair, will try another pass ;)

hehe :)

Quoted 11 lines

[surya@bluegenie tests]$ ./massive_intr 10 10

002961  00000121
002965  00000120
002964  00000121
002959  00000120
002956  00000121
002963  00000121
002960  00000121
002962  00000121
002958  00000122
002957  00000122
Quoted 3 lines

btw., other schedulers might work better with some more test-time: i'd suggest to use 60 seconds (./massive_intr 10 60) [or maybe more, using more threads] to see long-term fairness effects.

I tested CFS with massive_intr. I did long term, many CPUs, and many processes cases.

Test environment
================

 - kernel:         2.6.21-rc6-CFS
 - run time:       300 secs
 - # of CPU:       1 or 4

- # of processes: 200 or 800

Result
======

  +---------+-----------+-------+------+------+--------+
  |   # of  |   # of    | avg   | max  | min  |  stdev |
  |   CPUs  | processes | (*1)  | (*2) | (*3) |  (*4)  |
  +---------+-----------+-------+------+------+--------+
  | 1(i386) |           | 117.9 |  123 |  115 |    1.2 |
  +---------|       200 +-------+------+------+--------+
  |         |           | 750.2 |  767 |  735 |   10.6 |
  | 4(ia64) +-----------+-------+------+------+--------+
  |         |   800(*5) | 187.3 |  189 |  186 |    0.8 |
  +---------+-----------+-------+------+------+--------+

*1) average number of loops among all processes *2) maximum number of loops among all processes *3) minimum number of loops among all processes *4) standard deviation
*5) Its # of processes per CPU is equal to first test case.

Pretty good! CFS seems to be fair in any situation.

Satoru

#5 lore raw
Quoted 29 lines · Satoru Takeuchi
Quoted 3 lines

btw., other schedulers might work better with some more test-time:
i'd suggest to use 60 seconds (./massive_intr 10 60) [or maybe more, using more threads] to see long-term fairness effects.

I tested CFS with massive_intr. I did long term, many CPUs, and many processes cases.

Test environment
================

 - kernel:         2.6.21-rc6-CFS
 - run time:       300 secs
 - # of CPU:       1 or 4

- # of processes: 200 or 800

Result
======

  +---------+-----------+-------+------+------+--------+
  |   # of  |   # of    | avg   | max  | min  |  stdev |
  |   CPUs  | processes | (*1)  | (*2) | (*3) |  (*4)  |
  +---------+-----------+-------+------+------+--------+
  | 1(i386) |           | 117.9 |  123 |  115 |    1.2 |
  +---------|       200 +-------+------+------+--------+
  |         |           | 750.2 |  767 |  735 |   10.6 |
  | 4(ia64) +-----------+-------+------+------+--------+
  |         |   800(*5) | 187.3 |  189 |  186 |    0.8 |
  +---------+-----------+-------+------+------+--------+
Quoted 7 lines

*1) average number of loops among all processes *2) maximum number of loops among all processes *3) minimum number of loops among all processes *4) standard deviation
*5) Its # of processes per CPU is equal to first test case.

Pretty good! CFS seems to be fair in any situation.

thanks for testing this! Indeed the min-max values and standard deviation look all pretty healthy. (They in fact seem to be better than the other patch of mine against upstream that you tested, correct?)

[ And there's also another nice little detail in your feedback: CFS
actually builds, boots and works fine on ia64 too ;-) ]

Ingo

#6 lore raw

At Mon, 16 Apr 2007 10:47:25 +0200,

Quoted 39 lines · Ingo Molnar
Quoted 35 lines · Satoru Takeuchi
Quoted 3 lines

btw., other schedulers might work better with some more test-time:
i'd suggest to use 60 seconds (./massive_intr 10 60) [or maybe more, using more threads] to see long-term fairness effects.

I tested CFS with massive_intr. I did long term, many CPUs, and many processes cases.

Test environment
================

 - kernel:         2.6.21-rc6-CFS
 - run time:       300 secs
 - # of CPU:       1 or 4

- # of processes: 200 or 800

Result
======

  +---------+-----------+-------+------+------+--------+
  |   # of  |   # of    | avg   | max  | min  |  stdev |
  |   CPUs  | processes | (*1)  | (*2) | (*3) |  (*4)  |
  +---------+-----------+-------+------+------+--------+
  | 1(i386) |           | 117.9 |  123 |  115 |    1.2 |
  +---------|       200 +-------+------+------+--------+
  |         |           | 750.2 |  767 |  735 |   10.6 |
  | 4(ia64) +-----------+-------+------+------+--------+
  |         |   800(*5) | 187.3 |  189 |  186 |    0.8 |
  +---------+-----------+-------+------+------+--------+

*1) average number of loops among all processes *2) maximum number of loops among all processes *3) minimum number of loops among all processes *4) standard deviation
*5) Its # of processes per CPU is equal to first test case.

Quoted 5 lines
Quoted 1 line

Pretty good! CFS seems to be fair in any situation.

thanks for testing this! Indeed the min-max values and standard deviation look all pretty healthy. (They in fact seem to be better than the other patch of mine against upstream that you tested, correct?)

Is it?

patch:
http://lkml.org/lkml/2007/3/27/225

test result:
http://lkml.org/lkml/2007/3/31/41

If so, yes. That patch has some dispersion.

Quoted 3 lines

[ And there's also another nice little detail in your feedback: CFS
actually builds, boots and works fine on ia64 too ;-) ]

You are welcome. I can use larger machine by chance, and also tested there just now.

Test environment
================

 - kernel:         2.6.21-rc6-CFS
 - run time:       300 secs
 - # of CPU:       12

- # of processes: 200 or 2400

Result
===================================

  +----------+-----------+-------+------+------+--------+
  |   # of   |   # of    | avg   | max  | min  |  stdev |
  |   CPUs   | processes |       |      |      |        |
  +----------+-----------+-------+------+------+--------+
  |          |       200 |  2250 | 2348 | 2204 |     64 |
  | 12(ia64) +-----------+-------+------+------+--------+
  |          |      2400 | 187.5 |  197 |  176 |    4.3 |
  +----------+-----------+-------+------+------+--------+

Looks like good too.

BTW, I've a question. Actually this problem is fixed on CFS and DS. However they are mostly written from scratch and doeesn't suitable for stable version, for example 2.6.20.X. Can your other patch be compromise for stable version? Although that patch is not perfect, but I think it's preferable to leave it alone.

Satoru

#7 lore raw
Quoted 23 lines · Satoru Takeuchi

You are welcome. I can use larger machine by chance, and also tested there just now.

Test environment
================

 - kernel:         2.6.21-rc6-CFS
 - run time:       300 secs
 - # of CPU:       12

- # of processes: 200 or 2400

Result
===================================

  +----------+-----------+-------+------+------+--------+
  |   # of   |   # of    | avg   | max  | min  |  stdev |
  |   CPUs   | processes |       |      |      |        |
  +----------+-----------+-------+------+------+--------+
  |          |       200 |  2250 | 2348 | 2204 |     64 |
  | 12(ia64) +-----------+-------+------+------+--------+
  |          |      2400 | 187.5 |  197 |  176 |    4.3 |
  +----------+-----------+-------+------+------+--------+
Quoted 1 line

Looks like good too.

yeah. The spread between min and max is 11%, the spread between stddev and avg is 2.2%, which is quite OK for so many tasks.

Quoted 5 lines

BTW, I've a question. Actually this problem is fixed on CFS and DS. However they are mostly written from scratch and doeesn't suitable for stable version, for example 2.6.20.X. Can your other patch be compromise for stable version? Although that patch is not perfect, but I think it's preferable to leave it alone.

i'm afraid that small patch is not suitable for a general purpose Linux release (it hits interactivity way too much) - that's what this years-long struggle was about. But you could apply it to a special server-centric kernel.

Ingo