mbox series

[v3,0/8] sched/fair: rework the CFS load balance

Message ID 1568878421-12301-1-git-send-email-vincent.guittot@linaro.org
Headers show
Series sched/fair: rework the CFS load balance | expand

Message

Vincent Guittot Sept. 19, 2019, 7:33 a.m. UTC
Several wrong task placement have been raised with the current load
balance algorithm but their fixes are not always straight forward and
end up with using biased values to force migrations. A cleanup and rework
of the load balance will help to handle such UCs and enable to fine grain
the behavior of the scheduler for other cases.

Patch 1 has already been sent separately and only consolidate asym policy
in one place and help the review of the changes in load_balance.

Patch 2 renames the sum of h_nr_running in stats.

Patch 3 removes meaningless imbalance computation to make review of
patch 4 easier.

Patch 4 reworks load_balance algorithm and fixes some wrong task placement
but try to stay conservative.

Patch 5 add the sum of nr_running to monitor non cfs tasks and take that
into account when pulling tasks.

Patch 6 replaces runnable_load by load now that the signal is only used
when overloaded.

Patch 7 improves the spread of tasks at the 1st scheduling level.

Patch 8 uses utilization instead of load in all steps of misfit task
path.

Patch 9 replaces runnable_load_avg by load_avg in the wake up path.

Patch 10 optimizes find_idlest_group() that was using both runnable_load
and load. This has not been squashed with previous patch to ease the
review.

Some benchmarks results based on 8 iterations of each tests:
- small arm64 dual quad cores system

           tip/sched/core        w/ this patchset    improvement
schedpipe      54981 +/-0.36%        55459 +/-0.31%   (+0.97%)

hackbench
1 groups       0.906 +/-2.34%        0.906 +/-2.88%   (+0.06%)

- large arm64 2 nodes / 224 cores system

           tip/sched/core        w/ this patchset    improvement
schedpipe     125323 +/-0.98%       125624 +/-0.71%   (+0.24%)

hackbench -l (256000/#grp) -g #grp
1 groups      15.360 +/-1.76%       14.206 +/-1.40%   (+8.69%)
4 groups       5.822 +/-1.02%        5.508 +/-6.45%   (+5.38%)
16 groups      3.103 +/-0.80%        3.244 +/-0.77%   (-4.52%)
32 groups      2.892 +/-1.23%        2.850 +/-1.81%   (+1.47%)
64 groups      2.825 +/-1.51%        2.725 +/-1.51%   (+3.54%)
128 groups     3.149 +/-8.46%        3.053 +/-13.15%  (+3.06%)
256 groups     3.511 +/-8.49%        3.019 +/-1.71%  (+14.03%)

dbench
1 groups     329.677 +/-0.46%      329.771 +/-0.11%   (+0.03%)
4 groups     931.499 +/-0.79%      947.118 +/-0.94%   (+1.68%)
16 groups   1924.210 +/-0.89%     1947.849 +/-0.76%   (+1.23%)
32 groups   2350.646 +/-5.75%     2351.549 +/-6.33%   (+0.04%)
64 groups   2201.524 +/-3.35%     2192.749 +/-5.84%   (-0.40%)
128 groups  2206.858 +/-2.50%     2376.265 +/-7.44%   (+7.68%)
256 groups  1263.520 +/-3.34%     1633.143 +/-13.02% (+29.25%)

tip/sched/core sha1:
  0413d7f33e60 ('sched/uclamp: Always use 'enum uclamp_id' for clamp_id values')

Changes since v2:
- fix typo and reorder code
- some minor code fixes
- optimize the find_idles_group()

Not covered in this patchset:
- update find_idlest_group() to be more aligned with load_balance(). I didn't
  want to delay this version because of this update which is not ready yet
- Better detection of overloaded and fully busy state, especially for cases
  when nr_running > nr CPUs.

Vincent Guittot (8):
  sched/fair: clean up asym packing
  sched/fair: rename sum_nr_running to sum_h_nr_running
  sched/fair: remove meaningless imbalance calculation
  sched/fair: rework load_balance
  sched/fair: use rq->nr_running when balancing load
  sched/fair: use load instead of runnable load in load_balance
  sched/fair: evenly spread tasks when not overloaded
  sched/fair: use utilization to select misfit task
  sched/fair: use load instead of runnable load in wakeup path
  sched/fair: optimize find_idlest_group

 kernel/sched/fair.c | 805 +++++++++++++++++++++++++++-------------------------
 1 file changed, 417 insertions(+), 388 deletions(-)

-- 
2.7.4

Comments

Phil Auld Oct. 8, 2019, 2:32 p.m. UTC | #1
Hi Vincent,

On Thu, Sep 19, 2019 at 09:33:31AM +0200 Vincent Guittot wrote:
> Several wrong task placement have been raised with the current load

> balance algorithm but their fixes are not always straight forward and

> end up with using biased values to force migrations. A cleanup and rework

> of the load balance will help to handle such UCs and enable to fine grain

> the behavior of the scheduler for other cases.

> 

> Patch 1 has already been sent separately and only consolidate asym policy

> in one place and help the review of the changes in load_balance.

> 

> Patch 2 renames the sum of h_nr_running in stats.

> 

> Patch 3 removes meaningless imbalance computation to make review of

> patch 4 easier.

> 

> Patch 4 reworks load_balance algorithm and fixes some wrong task placement

> but try to stay conservative.

> 

> Patch 5 add the sum of nr_running to monitor non cfs tasks and take that

> into account when pulling tasks.

> 

> Patch 6 replaces runnable_load by load now that the signal is only used

> when overloaded.

> 

> Patch 7 improves the spread of tasks at the 1st scheduling level.

> 

> Patch 8 uses utilization instead of load in all steps of misfit task

> path.

> 

> Patch 9 replaces runnable_load_avg by load_avg in the wake up path.

> 

> Patch 10 optimizes find_idlest_group() that was using both runnable_load

> and load. This has not been squashed with previous patch to ease the

> review.

> 

> Some benchmarks results based on 8 iterations of each tests:

> - small arm64 dual quad cores system

> 

>            tip/sched/core        w/ this patchset    improvement

> schedpipe      54981 +/-0.36%        55459 +/-0.31%   (+0.97%)

> 

> hackbench

> 1 groups       0.906 +/-2.34%        0.906 +/-2.88%   (+0.06%)

> 

> - large arm64 2 nodes / 224 cores system

> 

>            tip/sched/core        w/ this patchset    improvement

> schedpipe     125323 +/-0.98%       125624 +/-0.71%   (+0.24%)

> 

> hackbench -l (256000/#grp) -g #grp

> 1 groups      15.360 +/-1.76%       14.206 +/-1.40%   (+8.69%)

> 4 groups       5.822 +/-1.02%        5.508 +/-6.45%   (+5.38%)

> 16 groups      3.103 +/-0.80%        3.244 +/-0.77%   (-4.52%)

> 32 groups      2.892 +/-1.23%        2.850 +/-1.81%   (+1.47%)

> 64 groups      2.825 +/-1.51%        2.725 +/-1.51%   (+3.54%)

> 128 groups     3.149 +/-8.46%        3.053 +/-13.15%  (+3.06%)

> 256 groups     3.511 +/-8.49%        3.019 +/-1.71%  (+14.03%)

> 

> dbench

> 1 groups     329.677 +/-0.46%      329.771 +/-0.11%   (+0.03%)

> 4 groups     931.499 +/-0.79%      947.118 +/-0.94%   (+1.68%)

> 16 groups   1924.210 +/-0.89%     1947.849 +/-0.76%   (+1.23%)

> 32 groups   2350.646 +/-5.75%     2351.549 +/-6.33%   (+0.04%)

> 64 groups   2201.524 +/-3.35%     2192.749 +/-5.84%   (-0.40%)

> 128 groups  2206.858 +/-2.50%     2376.265 +/-7.44%   (+7.68%)

> 256 groups  1263.520 +/-3.34%     1633.143 +/-13.02% (+29.25%)

> 

> tip/sched/core sha1:

>   0413d7f33e60 ('sched/uclamp: Always use 'enum uclamp_id' for clamp_id values')

> 

> Changes since v2:

> - fix typo and reorder code

> - some minor code fixes

> - optimize the find_idles_group()

> 

> Not covered in this patchset:

> - update find_idlest_group() to be more aligned with load_balance(). I didn't

>   want to delay this version because of this update which is not ready yet

> - Better detection of overloaded and fully busy state, especially for cases

>   when nr_running > nr CPUs.

> 

> Vincent Guittot (8):

>   sched/fair: clean up asym packing

>   sched/fair: rename sum_nr_running to sum_h_nr_running

>   sched/fair: remove meaningless imbalance calculation

>   sched/fair: rework load_balance

>   sched/fair: use rq->nr_running when balancing load

>   sched/fair: use load instead of runnable load in load_balance

>   sched/fair: evenly spread tasks when not overloaded

>   sched/fair: use utilization to select misfit task

>   sched/fair: use load instead of runnable load in wakeup path

>   sched/fair: optimize find_idlest_group

> 

>  kernel/sched/fair.c | 805 +++++++++++++++++++++++++++-------------------------

>  1 file changed, 417 insertions(+), 388 deletions(-)

> 

> -- 

> 2.7.4

> 


We've been testing v3 and for the most part everything looks good. The 
group imbalance issues are fixed on all of our test systems except one.

The one is an 8-node intel system with 160 cpus. I'll put the system 
details at the end. 

This shows the average number of benchmark threads running on each node
through the run. That is, not including the 2 stress jobs. The end 
results are a 4x slow down in the cgroup case versus not. The 152 and 
156 are the number of LU threads in the run. In all cases there are 2 
stress CPU threads running either in their own cgroups (GROUP) or 
everything is in one cgroup (NORMAL).  The normal case is pretty well 
balanced with only a few >= 20 and those that are are only a little 
over. In the GROUP cases things are not so good. There are some > 30 
for example, and others < 10.


lu.C.x_152_GROUP_1   17.52  16.86  17.90  18.52  20.00  19.00  22.00  20.19
lu.C.x_152_GROUP_2   15.70  15.04  15.65  15.72  23.30  28.98  20.09  17.52
lu.C.x_152_GROUP_3   27.72  32.79  22.89  22.62  11.01  12.90  12.14  9.93
lu.C.x_152_GROUP_4   18.13  18.87  18.40  17.87  18.80  19.93  20.40  19.60
lu.C.x_152_GROUP_5   24.14  26.46  20.92  21.43  14.70  16.05  15.14  13.16
lu.C.x_152_NORMAL_1  21.03  22.43  20.27  19.97  18.37  18.80  16.27  14.87
lu.C.x_152_NORMAL_2  19.24  18.29  18.41  17.41  19.71  19.00  20.29  19.65
lu.C.x_152_NORMAL_3  19.43  20.00  19.05  20.24  18.76  17.38  18.52  18.62
lu.C.x_152_NORMAL_4  17.19  18.25  17.81  18.69  20.44  19.75  20.12  19.75
lu.C.x_152_NORMAL_5  19.25  19.56  19.12  19.56  19.38  19.38  18.12  17.62

lu.C.x_156_GROUP_1   18.62  19.31  18.38  18.77  19.88  21.35  19.35  20.35
lu.C.x_156_GROUP_2   15.58  12.72  14.96  14.83  20.59  19.35  29.75  28.22
lu.C.x_156_GROUP_3   20.05  18.74  19.63  18.32  20.26  20.89  19.53  18.58
lu.C.x_156_GROUP_4   14.77  11.42  13.01  10.09  27.05  33.52  23.16  22.98
lu.C.x_156_GROUP_5   14.94  11.45  12.77  10.52  28.01  33.88  22.37  22.05
lu.C.x_156_NORMAL_1  20.00  20.58  18.47  18.68  19.47  19.74  19.42  19.63
lu.C.x_156_NORMAL_2  18.52  18.48  18.83  18.43  20.57  20.48  20.61  20.09
lu.C.x_156_NORMAL_3  20.27  20.00  20.05  21.18  19.55  19.00  18.59  17.36
lu.C.x_156_NORMAL_4  19.65  19.60  20.25  20.75  19.35  20.10  19.00  17.30
lu.C.x_156_NORMAL_5  19.79  19.67  20.62  22.42  18.42  18.00  17.67  19.42


From what I can see this was better but not perfect in v1.  It was closer and 
so the end results (LU reported times and op/s) were close enough. But looking 
closer at it there are still some issues. (NORMAL is comparable to above)


lu.C.x_152_GROUP_1   18.08  18.17  19.58  19.29  19.25  17.50  21.46  18.67
lu.C.x_152_GROUP_2   17.12  17.48  17.88  17.62  19.57  17.31  23.00  22.02
lu.C.x_152_GROUP_3   17.82  17.97  18.12  18.18  24.55  22.18  16.97  16.21
lu.C.x_152_GROUP_4   18.47  19.08  18.50  18.66  21.45  25.00  15.47  15.37
lu.C.x_152_GROUP_5   20.46  20.71  27.38  24.75  17.06  16.65  12.81  12.19

lu.C.x_156_GROUP_1   18.70  18.80  20.25  19.50  20.45  20.30  19.55  18.45
lu.C.x_156_GROUP_2   19.29  19.90  17.71  18.10  20.76  21.57  19.81  18.86
lu.C.x_156_GROUP_3   25.09  29.19  21.83  21.33  18.67  18.57  11.03  10.29
lu.C.x_156_GROUP_4   18.60  19.10  19.20  18.70  20.30  20.00  19.70  20.40
lu.C.x_156_GROUP_5   18.58  18.9   18.63  18.16  17.32  19.37  23.92  21.08

There is high variance so it may not be anythign specific between v1 and v3 here. 

The initial fixes I made for this issue did not exhibit this behavior. They 
would have had other issues dealing with overload cases though. In this case 
however there are only 154 or 158 threads on 160 CPUs so not overloaded. 

I'll try to get my hands on this system and poke into it. I just wanted to get 
your thoughts and let you know where we are. 



Thanks,
Phil


System details:

Architecture:        x86_64
CPU op-mode(s):      32-bit, 64-bit
Byte Order:          Little Endian
CPU(s):              160
On-line CPU(s) list: 0-159
Thread(s) per core:  2
Core(s) per socket:  10
Socket(s):           8
NUMA node(s):        8
Vendor ID:           GenuineIntel
CPU family:          6
Model:               47
Model name:          Intel(R) Xeon(R) CPU E7- 4870  @ 2.40GHz
Stepping:            2
CPU MHz:             1063.934
BogoMIPS:            4787.73
Virtualization:      VT-x
L1d cache:           32K
L1i cache:           32K
L2 cache:            256K
L3 cache:            30720K
NUMA node0 CPU(s):   0-9,80-89
NUMA node1 CPU(s):   10-19,90-99
NUMA node2 CPU(s):   20-29,100-109
NUMA node3 CPU(s):   30-39,110-119
NUMA node4 CPU(s):   40-49,120-129
NUMA node5 CPU(s):   50-59,130-139
NUMA node6 CPU(s):   60-69,140-149
NUMA node7 CPU(s):   70-79,150-159
Flags:               fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 cx16 xtpr pdcm pcid dca sse4_1 sse4_2 popcnt aes lahf_lm epb pti tpr_shadow vnmi flexpriority ept vpid dtherm ida arat

$ numactl --hardware
available: 8 nodes (0-7)
node 0 cpus: 0 1 2 3 4 5 6 7 8 9 80 81 82 83 84 85 86 87 88 89
node 0 size: 64177 MB
node 0 free: 60866 MB
node 1 cpus: 10 11 12 13 14 15 16 17 18 19 90 91 92 93 94 95 96 97 98 99
node 1 size: 64507 MB
node 1 free: 61167 MB
node 2 cpus: 20 21 22 23 24 25 26 27 28 29 100 101 102 103 104 105 106 107 108
109
node 2 size: 64507 MB
node 2 free: 61250 MB
node 3 cpus: 30 31 32 33 34 35 36 37 38 39 110 111 112 113 114 115 116 117 118
119
node 3 size: 64507 MB
node 3 free: 61327 MB
node 4 cpus: 40 41 42 43 44 45 46 47 48 49 120 121 122 123 124 125 126 127 128
129
node 4 size: 64507 MB
node 4 free: 60993 MB
node 5 cpus: 50 51 52 53 54 55 56 57 58 59 130 131 132 133 134 135 136 137 138
139
node 5 size: 64507 MB
node 5 free: 60892 MB
node 6 cpus: 60 61 62 63 64 65 66 67 68 69 140 141 142 143 144 145 146 147 148
149
node 6 size: 64507 MB
node 6 free: 61139 MB
node 7 cpus: 70 71 72 73 74 75 76 77 78 79 150 151 152 153 154 155 156 157 158
159
node 7 size: 64480 MB
node 7 free: 61188 MB
node distances:
node   0   1   2   3   4   5   6   7  
 0:  10  12  17  17  19  19  19  19  
 1:  12  10  17  17  19  19  19  19  
 2:  17  17  10  12  19  19  19  19  
 3:  17  17  12  10  19  19  19  19  
 4:  19  19  19  19  10  12  17  17  
 5:  19  19  19  19  12  10  17  17  
 6:  19  19  19  19  17  17  10  12  
 7:  19  19  19  19  17  17  12  10



--
Vincent Guittot Oct. 8, 2019, 3:53 p.m. UTC | #2
Hi Phil,

On Tue, 8 Oct 2019 at 16:33, Phil Auld <pauld@redhat.com> wrote:
>

> Hi Vincent,

>

> On Thu, Sep 19, 2019 at 09:33:31AM +0200 Vincent Guittot wrote:

> > Several wrong task placement have been raised with the current load

> > balance algorithm but their fixes are not always straight forward and

> > end up with using biased values to force migrations. A cleanup and rework

> > of the load balance will help to handle such UCs and enable to fine grain

> > the behavior of the scheduler for other cases.

> >


[...]

> >

>

> We've been testing v3 and for the most part everything looks good. The

> group imbalance issues are fixed on all of our test systems except one.

>

> The one is an 8-node intel system with 160 cpus. I'll put the system

> details at the end.

>

> This shows the average number of benchmark threads running on each node

> through the run. That is, not including the 2 stress jobs. The end

> results are a 4x slow down in the cgroup case versus not. The 152 and

> 156 are the number of LU threads in the run. In all cases there are 2

> stress CPU threads running either in their own cgroups (GROUP) or

> everything is in one cgroup (NORMAL).  The normal case is pretty well

> balanced with only a few >= 20 and those that are are only a little

> over. In the GROUP cases things are not so good. There are some > 30

> for example, and others < 10.

>

>

> lu.C.x_152_GROUP_1   17.52  16.86  17.90  18.52  20.00  19.00  22.00  20.19

> lu.C.x_152_GROUP_2   15.70  15.04  15.65  15.72  23.30  28.98  20.09  17.52

> lu.C.x_152_GROUP_3   27.72  32.79  22.89  22.62  11.01  12.90  12.14  9.93

> lu.C.x_152_GROUP_4   18.13  18.87  18.40  17.87  18.80  19.93  20.40  19.60

> lu.C.x_152_GROUP_5   24.14  26.46  20.92  21.43  14.70  16.05  15.14  13.16

> lu.C.x_152_NORMAL_1  21.03  22.43  20.27  19.97  18.37  18.80  16.27  14.87

> lu.C.x_152_NORMAL_2  19.24  18.29  18.41  17.41  19.71  19.00  20.29  19.65

> lu.C.x_152_NORMAL_3  19.43  20.00  19.05  20.24  18.76  17.38  18.52  18.62

> lu.C.x_152_NORMAL_4  17.19  18.25  17.81  18.69  20.44  19.75  20.12  19.75

> lu.C.x_152_NORMAL_5  19.25  19.56  19.12  19.56  19.38  19.38  18.12  17.62

>

> lu.C.x_156_GROUP_1   18.62  19.31  18.38  18.77  19.88  21.35  19.35  20.35

> lu.C.x_156_GROUP_2   15.58  12.72  14.96  14.83  20.59  19.35  29.75  28.22

> lu.C.x_156_GROUP_3   20.05  18.74  19.63  18.32  20.26  20.89  19.53  18.58

> lu.C.x_156_GROUP_4   14.77  11.42  13.01  10.09  27.05  33.52  23.16  22.98

> lu.C.x_156_GROUP_5   14.94  11.45  12.77  10.52  28.01  33.88  22.37  22.05

> lu.C.x_156_NORMAL_1  20.00  20.58  18.47  18.68  19.47  19.74  19.42  19.63

> lu.C.x_156_NORMAL_2  18.52  18.48  18.83  18.43  20.57  20.48  20.61  20.09

> lu.C.x_156_NORMAL_3  20.27  20.00  20.05  21.18  19.55  19.00  18.59  17.36

> lu.C.x_156_NORMAL_4  19.65  19.60  20.25  20.75  19.35  20.10  19.00  17.30

> lu.C.x_156_NORMAL_5  19.79  19.67  20.62  22.42  18.42  18.00  17.67  19.42

>

>

> From what I can see this was better but not perfect in v1.  It was closer and

> so the end results (LU reported times and op/s) were close enough. But looking

> closer at it there are still some issues. (NORMAL is comparable to above)

>

>

> lu.C.x_152_GROUP_1   18.08  18.17  19.58  19.29  19.25  17.50  21.46  18.67

> lu.C.x_152_GROUP_2   17.12  17.48  17.88  17.62  19.57  17.31  23.00  22.02

> lu.C.x_152_GROUP_3   17.82  17.97  18.12  18.18  24.55  22.18  16.97  16.21

> lu.C.x_152_GROUP_4   18.47  19.08  18.50  18.66  21.45  25.00  15.47  15.37

> lu.C.x_152_GROUP_5   20.46  20.71  27.38  24.75  17.06  16.65  12.81  12.19

>

> lu.C.x_156_GROUP_1   18.70  18.80  20.25  19.50  20.45  20.30  19.55  18.45

> lu.C.x_156_GROUP_2   19.29  19.90  17.71  18.10  20.76  21.57  19.81  18.86

> lu.C.x_156_GROUP_3   25.09  29.19  21.83  21.33  18.67  18.57  11.03  10.29

> lu.C.x_156_GROUP_4   18.60  19.10  19.20  18.70  20.30  20.00  19.70  20.40

> lu.C.x_156_GROUP_5   18.58  18.9   18.63  18.16  17.32  19.37  23.92  21.08

>

> There is high variance so it may not be anything specific between v1 and v3 here.


While preparing v4, I have noticed that I have probably oversimplified
the end of find_idlest_group() in patch "sched/fair: optimize
find_idlest_group" when it compares local vs the idlest other group.
Especially, there were a NUMA specific test that I removed in v3 and
re added in v4.

Then, I'm also preparing a full rework that find_idlest_group() which
will behave more closely to load_balance; I mean : collect statistics,
classify group then selects the idlest

What is the behavior of lu.C Thread ? are they waking up a lot  ?and
could trigger the slow wake path ?

>

> The initial fixes I made for this issue did not exhibit this behavior. They

> would have had other issues dealing with overload cases though. In this case

> however there are only 154 or 158 threads on 160 CPUs so not overloaded.

>

> I'll try to get my hands on this system and poke into it. I just wanted to get

> your thoughts and let you know where we are.


Thanks for testing

Vincent

>

>

>

> Thanks,

> Phil

>

>

> System details:

>

> Architecture:        x86_64

> CPU op-mode(s):      32-bit, 64-bit

> Byte Order:          Little Endian

> CPU(s):              160

> On-line CPU(s) list: 0-159

> Thread(s) per core:  2

> Core(s) per socket:  10

> Socket(s):           8

> NUMA node(s):        8

> Vendor ID:           GenuineIntel

> CPU family:          6

> Model:               47

> Model name:          Intel(R) Xeon(R) CPU E7- 4870  @ 2.40GHz

> Stepping:            2

> CPU MHz:             1063.934

> BogoMIPS:            4787.73

> Virtualization:      VT-x

> L1d cache:           32K

> L1i cache:           32K

> L2 cache:            256K

> L3 cache:            30720K

> NUMA node0 CPU(s):   0-9,80-89

> NUMA node1 CPU(s):   10-19,90-99

> NUMA node2 CPU(s):   20-29,100-109

> NUMA node3 CPU(s):   30-39,110-119

> NUMA node4 CPU(s):   40-49,120-129

> NUMA node5 CPU(s):   50-59,130-139

> NUMA node6 CPU(s):   60-69,140-149

> NUMA node7 CPU(s):   70-79,150-159

> Flags:               fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 cx16 xtpr pdcm pcid dca sse4_1 sse4_2 popcnt aes lahf_lm epb pti tpr_shadow vnmi flexpriority ept vpid dtherm ida arat

>

> $ numactl --hardware

> available: 8 nodes (0-7)

> node 0 cpus: 0 1 2 3 4 5 6 7 8 9 80 81 82 83 84 85 86 87 88 89

> node 0 size: 64177 MB

> node 0 free: 60866 MB

> node 1 cpus: 10 11 12 13 14 15 16 17 18 19 90 91 92 93 94 95 96 97 98 99

> node 1 size: 64507 MB

> node 1 free: 61167 MB

> node 2 cpus: 20 21 22 23 24 25 26 27 28 29 100 101 102 103 104 105 106 107 108

> 109

> node 2 size: 64507 MB

> node 2 free: 61250 MB

> node 3 cpus: 30 31 32 33 34 35 36 37 38 39 110 111 112 113 114 115 116 117 118

> 119

> node 3 size: 64507 MB

> node 3 free: 61327 MB

> node 4 cpus: 40 41 42 43 44 45 46 47 48 49 120 121 122 123 124 125 126 127 128

> 129

> node 4 size: 64507 MB

> node 4 free: 60993 MB

> node 5 cpus: 50 51 52 53 54 55 56 57 58 59 130 131 132 133 134 135 136 137 138

> 139

> node 5 size: 64507 MB

> node 5 free: 60892 MB

> node 6 cpus: 60 61 62 63 64 65 66 67 68 69 140 141 142 143 144 145 146 147 148

> 149

> node 6 size: 64507 MB

> node 6 free: 61139 MB

> node 7 cpus: 70 71 72 73 74 75 76 77 78 79 150 151 152 153 154 155 156 157 158

> 159

> node 7 size: 64480 MB

> node 7 free: 61188 MB

> node distances:

> node   0   1   2   3   4   5   6   7

>  0:  10  12  17  17  19  19  19  19

>  1:  12  10  17  17  19  19  19  19

>  2:  17  17  10  12  19  19  19  19

>  3:  17  17  12  10  19  19  19  19

>  4:  19  19  19  19  10  12  17  17

>  5:  19  19  19  19  12  10  17  17

>  6:  19  19  19  19  17  17  10  12

>  7:  19  19  19  19  17  17  12  10

>

>

>

> --
Phil Auld Oct. 9, 2019, 7:33 p.m. UTC | #3
On Tue, Oct 08, 2019 at 05:53:11PM +0200 Vincent Guittot wrote:
> Hi Phil,

>  


...

> While preparing v4, I have noticed that I have probably oversimplified

> the end of find_idlest_group() in patch "sched/fair: optimize

> find_idlest_group" when it compares local vs the idlest other group.

> Especially, there were a NUMA specific test that I removed in v3 and

> re added in v4.

> 

> Then, I'm also preparing a full rework that find_idlest_group() which

> will behave more closely to load_balance; I mean : collect statistics,

> classify group then selects the idlest

> 


Okay, I'll watch for V4 and restest. It really seems to be limited to 
the 8-node system. None of the other systems are showing this.


> What is the behavior of lu.C Thread ? are they waking up a lot  ?and

> could trigger the slow wake path ?


Yes, probably a fair bit of waking. It's an interative equation solving
code. It's fairly CPU intensive but requires communication for dependent
calculations.  That's part of why having them mis-balanced causes such a 
large slow down. I think at times everyone else waits for the slow guys.
 

> 

> >

> > The initial fixes I made for this issue did not exhibit this behavior. They

> > would have had other issues dealing with overload cases though. In this case

> > however there are only 154 or 158 threads on 160 CPUs so not overloaded.

> >

> > I'll try to get my hands on this system and poke into it. I just wanted to get

> > your thoughts and let you know where we are.

> 

> Thanks for testing

>


Sure!


Thanks,
Phil



 
> Vincent

> 

> >

> >

> >

> > Thanks,

> > Phil

> >

> >

> > System details:

> >

> > Architecture:        x86_64

> > CPU op-mode(s):      32-bit, 64-bit

> > Byte Order:          Little Endian

> > CPU(s):              160

> > On-line CPU(s) list: 0-159

> > Thread(s) per core:  2

> > Core(s) per socket:  10

> > Socket(s):           8

> > NUMA node(s):        8

> > Vendor ID:           GenuineIntel

> > CPU family:          6

> > Model:               47

> > Model name:          Intel(R) Xeon(R) CPU E7- 4870  @ 2.40GHz

> > Stepping:            2

> > CPU MHz:             1063.934

> > BogoMIPS:            4787.73

> > Virtualization:      VT-x

> > L1d cache:           32K

> > L1i cache:           32K

> > L2 cache:            256K

> > L3 cache:            30720K

> > NUMA node0 CPU(s):   0-9,80-89

> > NUMA node1 CPU(s):   10-19,90-99

> > NUMA node2 CPU(s):   20-29,100-109

> > NUMA node3 CPU(s):   30-39,110-119

> > NUMA node4 CPU(s):   40-49,120-129

> > NUMA node5 CPU(s):   50-59,130-139

> > NUMA node6 CPU(s):   60-69,140-149

> > NUMA node7 CPU(s):   70-79,150-159

> > Flags:               fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 cx16 xtpr pdcm pcid dca sse4_1 sse4_2 popcnt aes lahf_lm epb pti tpr_shadow vnmi flexpriority ept vpid dtherm ida arat

> >

> > $ numactl --hardware

> > available: 8 nodes (0-7)

> > node 0 cpus: 0 1 2 3 4 5 6 7 8 9 80 81 82 83 84 85 86 87 88 89

> > node 0 size: 64177 MB

> > node 0 free: 60866 MB

> > node 1 cpus: 10 11 12 13 14 15 16 17 18 19 90 91 92 93 94 95 96 97 98 99

> > node 1 size: 64507 MB

> > node 1 free: 61167 MB

> > node 2 cpus: 20 21 22 23 24 25 26 27 28 29 100 101 102 103 104 105 106 107 108

> > 109

> > node 2 size: 64507 MB

> > node 2 free: 61250 MB

> > node 3 cpus: 30 31 32 33 34 35 36 37 38 39 110 111 112 113 114 115 116 117 118

> > 119

> > node 3 size: 64507 MB

> > node 3 free: 61327 MB

> > node 4 cpus: 40 41 42 43 44 45 46 47 48 49 120 121 122 123 124 125 126 127 128

> > 129

> > node 4 size: 64507 MB

> > node 4 free: 60993 MB

> > node 5 cpus: 50 51 52 53 54 55 56 57 58 59 130 131 132 133 134 135 136 137 138

> > 139

> > node 5 size: 64507 MB

> > node 5 free: 60892 MB

> > node 6 cpus: 60 61 62 63 64 65 66 67 68 69 140 141 142 143 144 145 146 147 148

> > 149

> > node 6 size: 64507 MB

> > node 6 free: 61139 MB

> > node 7 cpus: 70 71 72 73 74 75 76 77 78 79 150 151 152 153 154 155 156 157 158

> > 159

> > node 7 size: 64480 MB

> > node 7 free: 61188 MB

> > node distances:

> > node   0   1   2   3   4   5   6   7

> >  0:  10  12  17  17  19  19  19  19

> >  1:  12  10  17  17  19  19  19  19

> >  2:  17  17  10  12  19  19  19  19

> >  3:  17  17  12  10  19  19  19  19

> >  4:  19  19  19  19  10  12  17  17

> >  5:  19  19  19  19  12  10  17  17

> >  6:  19  19  19  19  17  17  10  12

> >  7:  19  19  19  19  17  17  12  10

> >

> >

> >

> > --


--
Vincent Guittot Oct. 10, 2019, 8:20 a.m. UTC | #4
On Wed, 9 Oct 2019 at 21:33, Phil Auld <pauld@redhat.com> wrote:
>

> On Tue, Oct 08, 2019 at 05:53:11PM +0200 Vincent Guittot wrote:

> > Hi Phil,

> >

>

> ...

>

> > While preparing v4, I have noticed that I have probably oversimplified

> > the end of find_idlest_group() in patch "sched/fair: optimize

> > find_idlest_group" when it compares local vs the idlest other group.

> > Especially, there were a NUMA specific test that I removed in v3 and

> > re added in v4.

> >

> > Then, I'm also preparing a full rework that find_idlest_group() which

> > will behave more closely to load_balance; I mean : collect statistics,

> > classify group then selects the idlest

> >

>

> Okay, I'll watch for V4 and restest. It really seems to be limited to

> the 8-node system. None of the other systems are showing this.


Thanks for the information. This system might generate statistics at
boundaries between 2 behaviors and task placement misbehave

>

>

> > What is the behavior of lu.C Thread ? are they waking up a lot  ?and

> > could trigger the slow wake path ?

>

> Yes, probably a fair bit of waking. It's an interative equation solving

> code. It's fairly CPU intensive but requires communication for dependent

> calculations.  That's part of why having them mis-balanced causes such a

> large slow down. I think at times everyone else waits for the slow guys.

>

>

> >

> > >

> > > The initial fixes I made for this issue did not exhibit this behavior. They

> > > would have had other issues dealing with overload cases though. In this case

> > > however there are only 154 or 158 threads on 160 CPUs so not overloaded.

> > >

> > > I'll try to get my hands on this system and poke into it. I just wanted to get

> > > your thoughts and let you know where we are.

> >

> > Thanks for testing

> >

>

> Sure!

>

>

> Thanks,

> Phil

>

>

>

>

> > Vincent

> >

> > >

> > >

> > >

> > > Thanks,

> > > Phil

> > >

> > >

> > > System details:

> > >

> > > Architecture:        x86_64

> > > CPU op-mode(s):      32-bit, 64-bit

> > > Byte Order:          Little Endian

> > > CPU(s):              160

> > > On-line CPU(s) list: 0-159

> > > Thread(s) per core:  2

> > > Core(s) per socket:  10

> > > Socket(s):           8

> > > NUMA node(s):        8

> > > Vendor ID:           GenuineIntel

> > > CPU family:          6

> > > Model:               47

> > > Model name:          Intel(R) Xeon(R) CPU E7- 4870  @ 2.40GHz

> > > Stepping:            2

> > > CPU MHz:             1063.934

> > > BogoMIPS:            4787.73

> > > Virtualization:      VT-x

> > > L1d cache:           32K

> > > L1i cache:           32K

> > > L2 cache:            256K

> > > L3 cache:            30720K

> > > NUMA node0 CPU(s):   0-9,80-89

> > > NUMA node1 CPU(s):   10-19,90-99

> > > NUMA node2 CPU(s):   20-29,100-109

> > > NUMA node3 CPU(s):   30-39,110-119

> > > NUMA node4 CPU(s):   40-49,120-129

> > > NUMA node5 CPU(s):   50-59,130-139

> > > NUMA node6 CPU(s):   60-69,140-149

> > > NUMA node7 CPU(s):   70-79,150-159

> > > Flags:               fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 cx16 xtpr pdcm pcid dca sse4_1 sse4_2 popcnt aes lahf_lm epb pti tpr_shadow vnmi flexpriority ept vpid dtherm ida arat

> > >

> > > $ numactl --hardware

> > > available: 8 nodes (0-7)

> > > node 0 cpus: 0 1 2 3 4 5 6 7 8 9 80 81 82 83 84 85 86 87 88 89

> > > node 0 size: 64177 MB

> > > node 0 free: 60866 MB

> > > node 1 cpus: 10 11 12 13 14 15 16 17 18 19 90 91 92 93 94 95 96 97 98 99

> > > node 1 size: 64507 MB

> > > node 1 free: 61167 MB

> > > node 2 cpus: 20 21 22 23 24 25 26 27 28 29 100 101 102 103 104 105 106 107 108

> > > 109

> > > node 2 size: 64507 MB

> > > node 2 free: 61250 MB

> > > node 3 cpus: 30 31 32 33 34 35 36 37 38 39 110 111 112 113 114 115 116 117 118

> > > 119

> > > node 3 size: 64507 MB

> > > node 3 free: 61327 MB

> > > node 4 cpus: 40 41 42 43 44 45 46 47 48 49 120 121 122 123 124 125 126 127 128

> > > 129

> > > node 4 size: 64507 MB

> > > node 4 free: 60993 MB

> > > node 5 cpus: 50 51 52 53 54 55 56 57 58 59 130 131 132 133 134 135 136 137 138

> > > 139

> > > node 5 size: 64507 MB

> > > node 5 free: 60892 MB

> > > node 6 cpus: 60 61 62 63 64 65 66 67 68 69 140 141 142 143 144 145 146 147 148

> > > 149

> > > node 6 size: 64507 MB

> > > node 6 free: 61139 MB

> > > node 7 cpus: 70 71 72 73 74 75 76 77 78 79 150 151 152 153 154 155 156 157 158

> > > 159

> > > node 7 size: 64480 MB

> > > node 7 free: 61188 MB

> > > node distances:

> > > node   0   1   2   3   4   5   6   7

> > >  0:  10  12  17  17  19  19  19  19

> > >  1:  12  10  17  17  19  19  19  19

> > >  2:  17  17  10  12  19  19  19  19

> > >  3:  17  17  12  10  19  19  19  19

> > >  4:  19  19  19  19  10  12  17  17

> > >  5:  19  19  19  19  12  10  17  17

> > >  6:  19  19  19  19  17  17  10  12

> > >  7:  19  19  19  19  17  17  12  10

> > >

> > >

> > >

> > > --

>

> --
Parth Shah Oct. 16, 2019, 7:21 a.m. UTC | #5
On 9/19/19 1:03 PM, Vincent Guittot wrote:
> Several wrong task placement have been raised with the current load

> balance algorithm but their fixes are not always straight forward and

> end up with using biased values to force migrations. A cleanup and rework

> of the load balance will help to handle such UCs and enable to fine grain

> the behavior of the scheduler for other cases.

> 

> Patch 1 has already been sent separately and only consolidate asym policy

> in one place and help the review of the changes in load_balance.

> 

> Patch 2 renames the sum of h_nr_running in stats.

> 

> Patch 3 removes meaningless imbalance computation to make review of

> patch 4 easier.

> 

> Patch 4 reworks load_balance algorithm and fixes some wrong task placement

> but try to stay conservative.

> 

> Patch 5 add the sum of nr_running to monitor non cfs tasks and take that

> into account when pulling tasks.

> 

> Patch 6 replaces runnable_load by load now that the signal is only used

> when overloaded.

> 

> Patch 7 improves the spread of tasks at the 1st scheduling level.

> 

> Patch 8 uses utilization instead of load in all steps of misfit task

> path.

> 

> Patch 9 replaces runnable_load_avg by load_avg in the wake up path.

> 

> Patch 10 optimizes find_idlest_group() that was using both runnable_load

> and load. This has not been squashed with previous patch to ease the

> review.

> 

> Some benchmarks results based on 8 iterations of each tests:

> - small arm64 dual quad cores system

> 

>            tip/sched/core        w/ this patchset    improvement

> schedpipe      54981 +/-0.36%        55459 +/-0.31%   (+0.97%)

> 

> hackbench

> 1 groups       0.906 +/-2.34%        0.906 +/-2.88%   (+0.06%)

> 

> - large arm64 2 nodes / 224 cores system

> 

>            tip/sched/core        w/ this patchset    improvement

> schedpipe     125323 +/-0.98%       125624 +/-0.71%   (+0.24%)

> 

> hackbench -l (256000/#grp) -g #grp

> 1 groups      15.360 +/-1.76%       14.206 +/-1.40%   (+8.69%)

> 4 groups       5.822 +/-1.02%        5.508 +/-6.45%   (+5.38%)

> 16 groups      3.103 +/-0.80%        3.244 +/-0.77%   (-4.52%)

> 32 groups      2.892 +/-1.23%        2.850 +/-1.81%   (+1.47%)

> 64 groups      2.825 +/-1.51%        2.725 +/-1.51%   (+3.54%)

> 128 groups     3.149 +/-8.46%        3.053 +/-13.15%  (+3.06%)

> 256 groups     3.511 +/-8.49%        3.019 +/-1.71%  (+14.03%)

> 

> dbench

> 1 groups     329.677 +/-0.46%      329.771 +/-0.11%   (+0.03%)

> 4 groups     931.499 +/-0.79%      947.118 +/-0.94%   (+1.68%)

> 16 groups   1924.210 +/-0.89%     1947.849 +/-0.76%   (+1.23%)

> 32 groups   2350.646 +/-5.75%     2351.549 +/-6.33%   (+0.04%)

> 64 groups   2201.524 +/-3.35%     2192.749 +/-5.84%   (-0.40%)

> 128 groups  2206.858 +/-2.50%     2376.265 +/-7.44%   (+7.68%)

> 256 groups  1263.520 +/-3.34%     1633.143 +/-13.02% (+29.25%)

> 

> tip/sched/core sha1:

>   0413d7f33e60 ('sched/uclamp: Always use 'enum uclamp_id' for clamp_id values')

> [...]


I am quietly impressed with this patch series as it makes easy to
understand the behavior of the load balancer just by looking at the code.

I have tested v3 on IBM POWER9 system with following configuration:
- CPU(s):              176
- Thread(s) per core:  4
- Core(s) per socket:  22
- Socket(s):           2
- Model name:          POWER9, altivec supported
- NUMA node0 CPU(s):   0-87
- NUMA node8 CPU(s):   88-175

I see results in par with the baseline (tip/sched/core) with most of my
testings.

hackbench
=========
hackbench -l (256000/#grp) -g #grp (lower is better):
+--------+--------------------+-------------------+------------------+
| groups |     w/ patches     |     Baseline      | Performance gain |
+--------+--------------------+-------------------+------------------+
|      1 | 14.948 (+/- 0.10)  | 15.13 (+/- 0.47 ) | +1.20            |
|      4 | 5.938 (+/- 0.034)  | 6.085 (+/- 0.07)  | +2.4             |
|      8 | 6.594 (+/- 0.072)  | 6.223 (+/- 0.03)  | -5.9             |
|     16 | 5.916 (+/- 0.05)   | 5.559 (+/- 0.00)  | -6.4             |
|     32 | 5.288 (+/- 0.034)  | 5.23 (+/- 0.01)   | -1.1             |
|     64 | 5.147 (+/- 0.036)  | 5.193 (+/- 0.09)  | +0.8             |
|    128 | 5.368 (+/- 0.0245) | 5.446 (+/- 0.04)  | +1.4             |
|    256 | 5.637 (+/- 0.088)  | 5.596 (+/- 0.07)  | -0.7             |
|    512 | 5.78 (+/- 0.0637)  | 5.934 (+/- 0.06)  | +2.5             |
+--------+--------------------+-------------------+------------------+


dbench
========
dbench <grp> (Throughput: Higher is better):
+---------+---------------------+-----------------------+----------+
| groups  |     w/ patches      |        baseline       |    gain  |
+---------+---------------------+-----------------------+----------+
|       1 |    12.6419(+/-0.58) |    12.6511 (+/-0.277) |   -0.00  |
|       4 |    23.7712(+/-2.22) |    21.8526 (+/-0.844) |   +8.7   |
|       8 |    40.1333(+/-0.85) |    37.0623 (+/-3.283) |   +8.2   |
|      16 |   60.5529(+/-2.35)  |    60.0972 (+/-9.655) |   +0.7   |
|      32 |   98.2194(+/-1.69)  |    87.6701 (+/-10.72) |   +12.0  |
|      64 |   150.733(+/-9.91)  |    109.782 (+/-0.503) |   +37.3  |
|     128 |  173.443(+/-22.4)   |    130.006 (+/-21.84) |   +33.4  |
|     256 |  121.011(+/-15.2)   |    120.603 (+/-11.82) |   +0.3   |
|     512 |  10.9889(+/-0.39)   |    12.5518 (+/-1.030) |   -12    |
+---------+---------------------+-----------------------+----------+



I am happy with the results as it turns to be beneficial in most cases.
Still I will be doing testing for different scenarios and workloads.
BTW do you have any specific test case which might show different behavior
for SMT-4/8 systems with these patch set?


Thanks,
Parth
Vincent Guittot Oct. 16, 2019, 11:51 a.m. UTC | #6
Hi Parth,

On Wed, 16 Oct 2019 at 09:21, Parth Shah <parth@linux.ibm.com> wrote:
>

>

>

> On 9/19/19 1:03 PM, Vincent Guittot wrote:

> > Several wrong task placement have been raised with the current load

> > balance algorithm but their fixes are not always straight forward and

> > end up with using biased values to force migrations. A cleanup and rework

> > of the load balance will help to handle such UCs and enable to fine grain

> > the behavior of the scheduler for other cases.

> >

> > Patch 1 has already been sent separately and only consolidate asym policy

> > in one place and help the review of the changes in load_balance.

> >

> > Patch 2 renames the sum of h_nr_running in stats.

> >

> > Patch 3 removes meaningless imbalance computation to make review of

> > patch 4 easier.

> >

> > Patch 4 reworks load_balance algorithm and fixes some wrong task placement

> > but try to stay conservative.

> >

> > Patch 5 add the sum of nr_running to monitor non cfs tasks and take that

> > into account when pulling tasks.

> >

> > Patch 6 replaces runnable_load by load now that the signal is only used

> > when overloaded.

> >

> > Patch 7 improves the spread of tasks at the 1st scheduling level.

> >

> > Patch 8 uses utilization instead of load in all steps of misfit task

> > path.

> >

> > Patch 9 replaces runnable_load_avg by load_avg in the wake up path.

> >

> > Patch 10 optimizes find_idlest_group() that was using both runnable_load

> > and load. This has not been squashed with previous patch to ease the

> > review.

> >

> > Some benchmarks results based on 8 iterations of each tests:

> > - small arm64 dual quad cores system

> >

> >            tip/sched/core        w/ this patchset    improvement

> > schedpipe      54981 +/-0.36%        55459 +/-0.31%   (+0.97%)

> >

> > hackbench

> > 1 groups       0.906 +/-2.34%        0.906 +/-2.88%   (+0.06%)

> >

> > - large arm64 2 nodes / 224 cores system

> >

> >            tip/sched/core        w/ this patchset    improvement

> > schedpipe     125323 +/-0.98%       125624 +/-0.71%   (+0.24%)

> >

> > hackbench -l (256000/#grp) -g #grp

> > 1 groups      15.360 +/-1.76%       14.206 +/-1.40%   (+8.69%)

> > 4 groups       5.822 +/-1.02%        5.508 +/-6.45%   (+5.38%)

> > 16 groups      3.103 +/-0.80%        3.244 +/-0.77%   (-4.52%)

> > 32 groups      2.892 +/-1.23%        2.850 +/-1.81%   (+1.47%)

> > 64 groups      2.825 +/-1.51%        2.725 +/-1.51%   (+3.54%)

> > 128 groups     3.149 +/-8.46%        3.053 +/-13.15%  (+3.06%)

> > 256 groups     3.511 +/-8.49%        3.019 +/-1.71%  (+14.03%)

> >

> > dbench

> > 1 groups     329.677 +/-0.46%      329.771 +/-0.11%   (+0.03%)

> > 4 groups     931.499 +/-0.79%      947.118 +/-0.94%   (+1.68%)

> > 16 groups   1924.210 +/-0.89%     1947.849 +/-0.76%   (+1.23%)

> > 32 groups   2350.646 +/-5.75%     2351.549 +/-6.33%   (+0.04%)

> > 64 groups   2201.524 +/-3.35%     2192.749 +/-5.84%   (-0.40%)

> > 128 groups  2206.858 +/-2.50%     2376.265 +/-7.44%   (+7.68%)

> > 256 groups  1263.520 +/-3.34%     1633.143 +/-13.02% (+29.25%)

> >

> > tip/sched/core sha1:

> >   0413d7f33e60 ('sched/uclamp: Always use 'enum uclamp_id' for clamp_id values')

> > [...]

>

> I am quietly impressed with this patch series as it makes easy to

> understand the behavior of the load balancer just by looking at the code.


Thanks

>

> I have tested v3 on IBM POWER9 system with following configuration:

> - CPU(s):              176

> - Thread(s) per core:  4

> - Core(s) per socket:  22

> - Socket(s):           2

> - Model name:          POWER9, altivec supported

> - NUMA node0 CPU(s):   0-87

> - NUMA node8 CPU(s):   88-175

>

> I see results in par with the baseline (tip/sched/core) with most of my

> testings.

>

> hackbench

> =========

> hackbench -l (256000/#grp) -g #grp (lower is better):

> +--------+--------------------+-------------------+------------------+

> | groups |     w/ patches     |     Baseline      | Performance gain |

> +--------+--------------------+-------------------+------------------+

> |      1 | 14.948 (+/- 0.10)  | 15.13 (+/- 0.47 ) | +1.20            |

> |      4 | 5.938 (+/- 0.034)  | 6.085 (+/- 0.07)  | +2.4             |

> |      8 | 6.594 (+/- 0.072)  | 6.223 (+/- 0.03)  | -5.9             |

> |     16 | 5.916 (+/- 0.05)   | 5.559 (+/- 0.00)  | -6.4             |

> |     32 | 5.288 (+/- 0.034)  | 5.23 (+/- 0.01)   | -1.1             |

> |     64 | 5.147 (+/- 0.036)  | 5.193 (+/- 0.09)  | +0.8             |

> |    128 | 5.368 (+/- 0.0245) | 5.446 (+/- 0.04)  | +1.4             |

> |    256 | 5.637 (+/- 0.088)  | 5.596 (+/- 0.07)  | -0.7             |

> |    512 | 5.78 (+/- 0.0637)  | 5.934 (+/- 0.06)  | +2.5             |

> +--------+--------------------+-------------------+------------------+

>

>

> dbench

> ========

> dbench <grp> (Throughput: Higher is better):

> +---------+---------------------+-----------------------+----------+

> | groups  |     w/ patches      |        baseline       |    gain  |

> +---------+---------------------+-----------------------+----------+

> |       1 |    12.6419(+/-0.58) |    12.6511 (+/-0.277) |   -0.00  |

> |       4 |    23.7712(+/-2.22) |    21.8526 (+/-0.844) |   +8.7   |

> |       8 |    40.1333(+/-0.85) |    37.0623 (+/-3.283) |   +8.2   |

> |      16 |   60.5529(+/-2.35)  |    60.0972 (+/-9.655) |   +0.7   |

> |      32 |   98.2194(+/-1.69)  |    87.6701 (+/-10.72) |   +12.0  |

> |      64 |   150.733(+/-9.91)  |    109.782 (+/-0.503) |   +37.3  |

> |     128 |  173.443(+/-22.4)   |    130.006 (+/-21.84) |   +33.4  |

> |     256 |  121.011(+/-15.2)   |    120.603 (+/-11.82) |   +0.3   |

> |     512 |  10.9889(+/-0.39)   |    12.5518 (+/-1.030) |   -12    |

> +---------+---------------------+-----------------------+----------+

>

>

>

> I am happy with the results as it turns to be beneficial in most cases.

> Still I will be doing testing for different scenarios and workloads.


Thanks for the tests results

> BTW do you have any specific test case which might show different behavior

> for SMT-4/8 systems with these patch set?


No, I don't have any specific tests case in mind but it should better
spread tasks in cores for middle loaded UCs with tasks with different
weight either beacuse different nice prio or because of cgroup. The
load_balance doesn't use load in this case unlike current
implementation.

Thanks
Vincent

>

>

> Thanks,

> Parth

>