lkml.org 
[lkml]   [2012]   [May]   [27]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
SubjectRe: [rfc][patch] select_idle_sibling() inducing bouncing on westmere
From
Date
On Sat, 2012-05-26 at 10:27 +0200, Mike Galbraith wrote: 
> Hohum, back to finding out what happened to cpufreq.

Answer: nothing.. in mainline.

I test performance habitually, so just never noticed how bad ondemand
sucks. In enterprise, I found the below, explaining why cores crank up
fine there, but not in mainline. Somebody thumped ondemand properly on
it's pointy head.

But, check out the numbers below this, and you can see just how horrible
bouncing is when you add governor latency _on top_ of it.

---
drivers/cpufreq/cpufreq_ondemand.c | 25 +++++++++++++++++++++++++
1 file changed, 25 insertions(+)
--- a/drivers/cpufreq/cpufreq_ondemand.c
+++ b/drivers/cpufreq/cpufreq_ondemand.c
@@ -37,6 +37,7 @@
#define MICRO_FREQUENCY_MIN_SAMPLE_RATE (10000)
#define MIN_FREQUENCY_UP_THRESHOLD (11)
#define MAX_FREQUENCY_UP_THRESHOLD (100)
+#define MAX_DEFAULT_SAMPLING_RATE (300 * 1000U)

/*
* The polling frequency of this governor depends on the capability of
@@ -733,6 +734,30 @@ static int cpufreq_governor_dbs(struct c
max(min_sampling_rate,
latency * LATENCY_MULTIPLIER);
dbs_tuners_ins.io_is_busy = should_io_be_busy();
+ /*
+ * Cut def_sampling rate to 300ms if it was above,
+ * still consider to not set it above latency
+ * transition * 100
+ */
+ if (dbs_tuners_ins.sampling_rate > MAX_DEFAULT_SAMPLING_RATE) {
+ dbs_tuners_ins.sampling_rate =
+ max(min_sampling_rate, MAX_DEFAULT_SAMPLING_RATE);
+ printk(KERN_INFO "CPUFREQ: ondemand sampling "
+ "rate set to %d ms\n",
+ dbs_tuners_ins.sampling_rate / 1000);
+ }
+ /*
+ * Be conservative in respect to performance.
+ * If an application calculates using two threads
+ * depending on each other, they will be run on several
+ * CPU cores resulting on 50% load on both.
+ * SLED might still want to prefer 80% up_threshold
+ * by default, but we cannot differ that here.
+ */
+ if (num_online_cpus() > 1)
+ dbs_tuners_ins.up_threshold =
+ DEF_FREQUENCY_UP_THRESHOLD / 2;
+
}
mutex_unlock(&dbs_mutex);


patches applied to both trees
patches/remove_irritating_plus.diff
patches/clockevents-Reinstate-the-per-cpu-tick-skew.patch
patches/sched-cgroups-Disallow-attaching-kthreadd
patches/sched-fix-task_groups-list
patches/sched-rt-fix-isolated-CPUs-leaving-root_task_group-indefinitely-throttled.patch
patches/sched-throttle-nohz.patch
patches/sched-domain-flags-proc-handler.patch
patches/sched-fix-Q6600.patch
patches/cpufreq_ondemand_performance_optimise_default_settings.patch

applied only to 3.4.0x
patches/sched-tweak-select_idle_sibling.patch

tbench 1
3.4.0 351 MB/sec ondemand
350 MB/sec
351 MB/sec
3.4.0x 428 MB/sec ondemand
432 MB/sec
425 MB/sec
vs 3.4.0 1.22

3.4.0 363 MB/sec performance
369 MB/sec
359 MB/sec

3.4.0x 432 MB/sec performance
430 MB/sec
427 MB/sec
vs 3.4.0 1.18

netperf TCP_RR 1 byte ping/pong (trans/sec)

governor ondemand
unbound bound
3.4.0 72851 128433
72347 127301
72512 127472

3.4.0x 128440 131979
128116 132413
128366 132004
vs 3.4.0 1.768 1.034
^^^^^ eek! (hm, why bound improvement?)
governor performance
3.4.0 105199 127140
104534 128786
104167 127920

3.4.0x 123451 132883
128702 132688
125653 133005
vs 3.4.0 1.203 1.038
(hm, why bound improvement?)

select_idle_sibling() becomes a proper throughput/latency trade on
Westmere as well, with only modest cost even for worst case load that
does at least a dinky bit of work (TCP_RR == 100% synchronous).

-Mike



\
 
 \ /
  Last update: 2012-05-27 12:01    [W:0.153 / U:0.372 seconds]
©2003-2014 Jasper Spaans. Advertise on this site