Re: [PATCH v7 0/8] Tunable sched_mc_power_savings=n

[Vaidyanathan Srinivasan <svaidy@linux.vnet.ibm.com>]

* Vaidyanathan Srinivasan <svaidy@linux.vnet.ibm.com> [2008-12-30 23:37:22]:

> * Ingo Molnar <mingo@elte.hu> [2008-12-30 07:21:39]:
> 
> > 
> > * Balbir Singh <balbir@linux.vnet.ibm.com> wrote:
> > 
> > > > > KERNBENCH Runs: make -j4 on a x86 8 core, dual socket quad core cpu
> > > > > package system
> > > > >
> > > > > SchedMC Run Time     Package Idle    Energy  Power
> > > > > 0       81.68        52.83% 54.71%  1.00x J 1.00y W
> > > > > 1       80.70        36.62% 70.11%  0.95x J 0.96y W
> > > > > 2       74.95        19.53% 85.92%  0.90x J 0.98y W
> > 
> > > > Your result is very interesting.
> > > > level 2 is more fast and efficient of power.
> > > > 
> > > > What's major contributor to use less time in level 2?
> > > > I think it's cache bounce is less time than old.
> > > > Is right ?
> > > 
> > > Yes, correct
> > 
> > yes, i too noticed that runtime improved so dramatically: +7.5% on 
> > kernbench is a _very_ big deal.
> > 
> > So i wanted to ask you to re-test whether this speedup is reproducible, 
> > and if yes, please check a few other workloads (for example sysbench on 
> > postgresql / mysqld) and send a patch that changes the 
> > sched_mc_power_savings default to POWERSAVINGS_BALANCE_WAKEUP (2).
> 
> The speedup for kernbench is reproducible.  I will post a more
> detailed test report on kernbench soon.  The power-performance benefit
> is for an under-utilised system (nr_cpus/2) run of kernbench which is
> very ideal to demonstrate the power savings feature.  I will also try
> sysbench and update results little later.  As Balbir mentioned, I am
> on vacation and traveling.  I will post more benchmark results as soon
> as possible.

Hi Ingo,

I ran my kernbench test setup over multiple iterations with several
thread counts.  The performance will almost max out at 8-10 threads of
kernbench since the system has total of 8 logical CPUs.

The power saving benefits is expected to be maximum around 30-50%
system utilisation (3-5 threads).

I have posted the normalised results in various tables below.
Apparently the sched_mc=2 settings is helping at 8 threads case also
due to better cache line sharing.  I will work with Balbir to see if
we can prove this assertion using performance counters.  I will start
looking at sysbench and try to get benchmark results and power/energy
data in similar configurations.

--Vaidy

All these tests were done on sched-tip at 
commit: fe235c6b2a4b9eb11909fe40249abcce67f7d45d 
uname -a is 2.6.28-rc8-tip-sv-05664-gfe235c6-dirty

Kernbench was run on source 2.6.28 while the previous tests used
2.6.25 kernel which had far less stuff to compile in defconfig :)

System configuration is x86 quad core dual socket system 
with 8 logical CPUs

Each result is an average of 5 iterations.

Kernbench threads = 2

SchedMC Run Time     Package Idle(%)    Energy(J)  Power(W)
0	217.37	     73.30 76.93	1.00	   1.00
1	204.68	     50.86 99.91	0.95	   1.01
2	204.49	     50.69 99.93	0.95	   1.01

Kernbench threads = 4

SchedMC Run Time     Package Idle(%)    Energy(J)  Power(W)
0	109.11	     51.99 54.91	1.00	   1.00
1	109.71	     35.86 71.88	0.96	   0.96
2	102.95	     25.02 82.47	0.92	   0.97

Kernbench threads = 6

SchedMC Run Time     Package Idle(%)    Energy(J)  Power(W)
0	74.77	     35.29 36.37	1.00	   1.00
1	74.54	     28.83 40.90	0.99	   0.99
2	73.38	     21.09 47.29	0.98	   1.00

Kernbench threads = 8

SchedMC Run Time     Package Idle(%)    Energy(J)  Power(W)
0	63.18	     26.23 28.67	1.00	   1.00
1	56.89	     19.85 22.38 	0.93	   1.02
2	55.27	     17.65 19.84	0.91	   1.03


Relative measures for sched_mc=1 compared to baseline sched_mc=0

Threads=>	2	4	6	8	10	12
KernbenchTime	0.94	1.01	1.00	0.90	0.93	0.95
Energy		0.95	0.96	0.99	0.93	0.95	0.97
Power		1.01	0.96	0.99	1.02	1.02	1.01

Relative measures for sched_mc=2 compared to baseline sched_mc=0

Threads=>	2	4	6	8	10	12
KernbenchTime	0.94	0.94	0.98	0.87	0.92	0.94
Energy		0.95	0.92	0.98	0.91	0.94	0.96
Power		1.01	0.97	1.00	1.03	1.02	1.02