RE: cpufreq_ondemand governor - problem

From: Starikovskiy, Alexey Y
Date: Sat Jul 15 2006 - 03:06:15 EST


Beside ondemand governor there is processor driver who should do
synchronization of frequencies over dependent CPUs (cores in your case).
If policy->cpus mask is set, then ondemand governor will choose minimum
idle time over dependent cores, and calculate load from it. If driver
does set policy-cpus mask, it's his job, or job of the processor itself
to do synchronization.

Hope that helps,
Alex

>-----Original Message-----
>From: art [mailto:art@xxxxxxxxxx]
>Sent: Saturday, July 15, 2006 2:42 AM
>To: linux-kernel@xxxxxxxxxxxxxxx
>Cc: Pallipadi, Venkatesh; Starikovskiy, Alexey Y;
>torvalds@xxxxxxxx; akpm@xxxxxxxx; mingo@xxxxxxx
>Subject: cpufreq_ondemand governor - problem
>
>problem:
>on dualcore AMD - if you use cpufreq_ondemand governor and
>your program is
>one_process/one_thread intensive one core is busy and second is doing
>nothing - governor is droping speed on both cores to lowest
>speed - slowing
>down busy core process - my dualcore-AMD do this i'm not shure
>if it is only
>AMD or INTEL problem too
>
>to test this set ondemand governor
>
># echo "ondemand" >
>/sys/devices/system/cpu/cpu0/cpufreq/scaling_governor
>
>now
>start in terminal-1
>
>#awk 'BEGIN {for(i=0;i<100000;i++)for(j=0;j<100000;j++);}'
>
>observe cpu speed and utilization
>core1 - utilization 100% speed lowest possible
>core2 - utilization 0% speed lowest possible
>
>now
>start in terminal-2
>
>#awk 'BEGIN {for(i=0;i<100000;i++)for(j=0;j<100000;j++);}'
>
>observe cpu speed and utilization
>core1 - utilization 100% speed max possible
>core2 - utilization 100% speed max possible
>
>now kill one awk
>
>observe cpu speed and utilization
>core1 - utilization 100% speed lowest possible
>core2 - utilization 0% speed lowest possible
>
>looks like cpufreq ondemand governor sets two frequency
>dependent cores to
>speed level ok for that one with lowest utilization slowing down
>process/thread working on other core. For now it is ok for independent
>multiprocessor bad for multicore-freq-dependent.
>
>
>temporary dirty patch works for me - your result my vary (for
>shure it will
>not work for multi-processor/dualcore - we need identify and
>pair cores to
>do same thing)
>
>
> --- cpufreq_ondemand.c-org 2006-07-05 23:09:49.000000000 -0500
>+++ cpufreq_ondemand.c 2006-07-14 15:50:56.000000000 -0500
>@@ -39,6 +39,7 @@
> * All times here are in uS.
> */
> static unsigned int def_sampling_rate;
>+static unsigned int load_max_core=0;
> #define MIN_SAMPLING_RATE_RATIO (2)
> /* for correct statistics, we need at least 10 ticks between
>each measure
>*/
> #define MIN_STAT_SAMPLING_RATE
>(MIN_SAMPLING_RATE_RATIO *
>jiffies_to_usecs(10))
>@@ -268,6 +269,8 @@ static void dbs_check_cpu(struct cpu_dbs
> idle_ticks = tmp_idle_ticks;
> }
> load = (100 * (total_ticks - idle_ticks)) / total_ticks;
>+ if (load_max_core > load)
>+ load = load_max_core;
>
> /* Check for frequency increase */
> if (load > dbs_tuners_ins.up_threshold) {
>@@ -297,6 +300,7 @@ static void dbs_check_cpu(struct cpu_dbs
>
> __cpufreq_driver_target(policy, freq_next,
>CPUFREQ_RELATION_L);
> }
>+load_max_core = 0;
> }
>
> static void do_dbs_timer(void *data)
>
>
> ---------------------------------------------------
>
>after this patch dualcore-AMD is working OK max speed for 100%
>utilization
>on core1 and 0% utilization on core2
>
>
>xboom
>art@xxxxxxxxxx
>
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/