Hi Vincent,
I conducted test runs of ebizzy on a Power8 box which had 48 cpus. 6 cores with SMT-8 to be precise. Its a single socket box. The results are as below.
On 05/23/2014 09:22 PM, Vincent Guittot wrote:
Part of this patchset was previously part of the larger tasks packing patchset [1]. I have splitted the latter in 3 different patchsets (at least) to make the thing easier. -configuration of sched_domain topology [2] -update and consolidation of cpu_power (this patchset) -tasks packing algorithm
SMT system is no more the only system that can have a CPUs with an original capacity that is different from the default value. We need to extend the use of cpu_power_orig to all kind of platform so the scheduler will have both the maximum capacity (cpu_power_orig/power_orig) and the current capacity (cpu_power/power) of CPUs and sched_groups. A new function arch_scale_cpu_power has been created and replace arch_scale_smt_power, which is SMT specifc in the computation of the capapcity of a CPU.
During load balance, the scheduler evaluates the number of tasks that a group of CPUs can handle. The current method assumes that tasks have a fix load of SCHED_LOAD_SCALE and CPUs have a default capacity of SCHED_POWER_SCALE. This assumption generates wrong decision by creating ghost cores and by removing real ones when the original capacity of CPUs is different from the default SCHED_POWER_SCALE.
Now that we have the original capacity of a CPUS and its activity/utilization, we can evaluate more accuratly the capacity of a group of CPUs.
This patchset mainly replaces the old capacity method by a new one and has kept the policy almost unchanged whereas we can certainly take advantage of this new statistic in several other places of the load balance.
TODO:
- align variable's and field's name with the renaming [3]
Tests results: I have put below results of 2 tests:
- hackbench -l 500 -s 4096
- scp of 100MB file on the platform
on a dual cortex-A7 hackbench scp tip/master 25.75s(+/-0.25) 5.16MB/s(+/-1.49)
- patches 1,2 25.89s(+/-0.31) 5.18MB/s(+/-1.45)
- patches 3-10 25.68s(+/-0.22) 7.00MB/s(+/-1.88)
- irq accounting 25.80s(+/-0.25) 8.06MB/s(+/-0.05)
on a quad cortex-A15 hackbench scp tip/master 15.69s(+/-0.16) 9.70MB/s(+/-0.04)
- patches 1,2 15.53s(+/-0.13) 9.72MB/s(+/-0.05)
- patches 3-10 15.56s(+/-0.22) 9.88MB/s(+/-0.05)
- irq accounting 15.99s(+/-0.08) 10.37MB/s(+/-0.03)
The improvement of scp bandwidth happens when tasks and irq are using different CPU which is a bit random without irq accounting config
N -> Number of threads of ebizzy
Each 'N' run was for 30 seconds with multiple iterations and averaging them.
N %change in number of records read after patching ------------------------------------------ 1 + 0.0038 4 -17.6429 8 -26.3989 12 -29.5070 16 -38.4842 20 -44.5747 24 -51.9792 28 -34.1863 32 -38.4029 38 -22.2490 42 -7.4843 47 -0.69676
Let me profile it and check where the cause of this degradation is.
Regards Preeti U Murthy