Re: [ck] Re: [PATCH] implement nice support across physical cpus on SMP

[Con Kolivas <kernel@kolivas.org>]

[-- Attachment #1: Type: text/plain, Size: 4218 bytes --]

On Mon, 9 May 2005 21:24, Markus Törnqvist wrote:
> I beg to differ with Mr. Carvalho's assesment with this patch;
> it works like a charm, and then some.
>
> The rest of the message is just my analysis of the situation
> ran on a Dell PowerEdge 2850, dual hyperthread Xeon EM64Ts, with
> Debian Pure64 Sarge installed.

Thanks for feedback.

>   PID USER      PR  NI  VIRT  SHR S %CPU %MEM    TIME+  #C COMMAND
>  3917 mjt       26   0  5660  936 R 99.9  0.0   3:37.61  0 load_base.sh
>  3918 mjt       25   0  5660  936 R 99.9  0.0   3:37.60  3 load_base.sh
>  3916 mjt       26   0  5660  936 R 52.7  0.0   2:02.37  2 load_base.sh
>  3919 mjt       39  19  5660  936 R  7.0  0.0   0:13.80  1 load_base.sh
>  3920 mjt       39  19  5660  936 R  3.0  0.0   0:06.05  2 load_base.sh
>
> top - 11:09:24 up 15:30,  2 users,  load average: 4.99, 3.55, 1.63
>   PID USER      PR  NI  VIRT  SHR S %CPU %MEM    TIME+  #C COMMAND
>  3917 mjt       25   0  5660  936 R 99.6  0.0   6:11.35  0 load_base.sh
>  3918 mjt       24   0  5660  936 R 99.6  0.0   6:11.34  3 load_base.sh
>  3916 mjt       39   0  5660  936 R 65.7  0.0   3:28.95  2 load_base.sh
>  3919 mjt       39  19  5660  936 R  7.0  0.0   0:23.54  1 load_base.sh
>  3920 mjt       39  19  5660  936 R  3.0  0.0   0:10.33  2 load_base.sh

These runs don't look absolutely "ideal" as one nice 19 task is bound to cpu1 
however since you're running hyperthreading it would seem the SMT nice code 
is keeping that under check anyway (0:23 vs 6:11)

> top - 11:12:32 up 15:33,  2 users,  load average: 4.99, 4.22, 2.24
>   PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  #C COMMAND
>  3917 mjt       35   0  5660 1136  936 R 99.9  0.0   9:28.56  0
> load_base.sh 3918 mjt       34   0  5660 1136  936 R 99.5  0.0   9:28.54  3
> load_base.sh 3916 mjt       35   0  5660 1136  936 R 61.7  0.0   5:19.77  2
> load_base.sh 3919 mjt       39  19  5660 1136  936 R  6.0  0.0   0:36.07  1
> load_base.sh 3920 mjt       39  19  5660 1136  936 R  3.0  0.0   0:15.82  2
> load_base.sh
>
> $ ./load.sh 7
> top - 11:13:49 up 15:35,  2 users,  load average: 5.17, 4.40, 2.45
>   PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  #C COMMAND
>  3952 mjt       29   0  5660 1140  936 R 99.9  0.0   0:33.53  2
> load_base.sh 3950 mjt       31   0  5660 1140  936 R 99.5  0.0   0:33.33  1
> load_base.sh 3953 mjt       30   0  5660 1140  936 R 55.7  0.0   0:16.82  3
> load_base.sh 3951 mjt       39   0  5660 1140  936 R 43.8  0.0   0:16.70  3
> load_base.sh 3949 mjt       39   0  5660 1140  936 R 23.9  0.0   0:13.18  0
> load_base.sh 3954 mjt       39  19  5660 1140  936 R  2.0  0.0   0:00.64  0
> load_base.sh 3955 mjt       39  19  5660 1140  936 R  2.0  0.0   0:00.64  0
> load_base.sh
>
> top - 11:14:53 up 15:36,  2 users,  load average: 6.38, 4.91, 2.76
>   PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  #C COMMAND
>  3950 mjt       23   0  5660 1140  936 R 99.9  0.0   1:39.67  1
> load_base.sh 3952 mjt       21   0  5660 1140  936 R 99.9  0.0   1:39.87  2
> load_base.sh 3951 mjt       39   0  5660 1140  936 R 52.7  0.0   0:49.91  3
> load_base.sh 3953 mjt       22   0  5660 1140  936 R 47.8  0.0   0:49.95  3
> load_base.sh 3949 mjt       39   0  5660 1140  936 R 43.8  0.0   0:38.70  0
> load_base.sh 3954 mjt       39  19  5660 1140  936 R  2.0  0.0   0:01.90  0
> load_base.sh 3955 mjt       39  19  5660 1140  936 R  2.0  0.0   0:01.90  0

These runs pretty much confirm what I found to happen. My test machine for 
this was also 4x. I can't see how the code would behave differently on 2x. 
Perhaps if I make the prio_bias multiplied instead of added to the cpu load 
it will be less affected by SCHED_LOAD_SCALE. The attached patch was 
confirmed during testing to also provide smp distribution according to nice 
on 4x. Carlos I know your machine is in production so you testing may not be 
easy for you. Please try this on top if you have time.

Cheers,
Con

---
This patch alters the effect priority bias has on busy rebalancing by 
multiplying the cpu load by the total priority instead of adding it.

Signed-off-by: Con Kolivas <kernel@kolivas.org>

[-- Attachment #2: alter_prio_bias.diff --]
[-- Type: text/x-diff, Size: 912 bytes --]

Index: linux-2.6.11-smpnice/kernel/sched.c
===================================================================
--- linux-2.6.11-smpnice.orig/kernel/sched.c	2005-05-07 23:25:15.000000000 +1000
+++ linux-2.6.11-smpnice/kernel/sched.c	2005-05-09 21:42:02.000000000 +1000
@@ -953,8 +953,8 @@ static inline unsigned long source_load(
 		 * If we are balancing busy runqueues the load is biased by
 		 * priority to create 'nice' support across cpus.
 		 */
-		cpu_load += rq->prio_bias;
-		load_now += rq->prio_bias;
+		cpu_load *= rq->prio_bias;
+		load_now *= rq->prio_bias;
 	}
 	return min(cpu_load, load_now);
 }
@@ -969,8 +969,8 @@ static inline unsigned long target_load(
 		load_now = rq->nr_running * SCHED_LOAD_SCALE;
 
 	if (idle == NOT_IDLE) {
-		cpu_load += rq->prio_bias;
-		load_now += rq->prio_bias;
+		cpu_load *= rq->prio_bias;
+		load_now *= rq->prio_bias;
 	}
 	return max(cpu_load, load_now);
 }