Re: [PATCH v2] sched: Initialize the vruntime of a new task when it is first enqueued

[Mukesh Kumar Chaurasiya <mchauras@linux.ibm.com>]

On Thu, 27 Jun 2024 21:33:59 +0800, Zhang Qiao wrote:
> When creating a new task, we initialize vruntime of the newly task at
> sched_cgroup_fork(). However, the timing of executing this action is too
> early and may not be accurate.
> 
> Because it uses current CPU to init the vruntime, but the new task
> actually runs on the cpu which be assigned at wake_up_new_task().
> 
> [...]

I tried this on a 104 cpu ppc64le Power9 system

The experiment was to stress the cpus with 416 threads(4 * nproc)
and then run small commands like date, time, ls etc

With bpf i collected the time it took for the tasks from waking up
to switching on to the cpu.

I collected 3600 samples for each

The results are as follows:

With patch

@: count 3601, average 4609830, total 16599997845

@c: 3709
@latency: 
[4K, 8K)               1 |                                                    |
[8K, 16K)              4 |                                                    |
[16K, 32K)          2270 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
[32K, 64K)            45 |@                                                   |
[64K, 128K)            3 |                                                    |
[128K, 256K)         141 |@@@                                                 |
[256K, 512K)          24 |                                                    |
[512K, 1M)             7 |                                                    |
[1M, 2M)               8 |                                                    |
[2M, 4M)              38 |                                                    |
[4M, 8M)             226 |@@@@@                                               |
[8M, 16M)            321 |@@@@@@@                                             |
[16M, 32M)           501 |@@@@@@@@@@@                                         |
[32M, 64M)            12 |                                                    |

# Without Patch

@: count 3600, average 12786599, total 46031757085

@c: 3608
@latency: 
[4K, 8K)               1 |                                                    |
[8K, 16K)              6 |                                                    |
[16K, 32K)           297 |@@@@@@@@@@@                                         |
[32K, 64K)            25 |                                                    |
[64K, 128K)            4 |                                                    |
[128K, 256K)         149 |@@@@@                                               |
[256K, 512K)          21 |                                                    |
[512K, 1M)             5 |                                                    |
[1M, 2M)              14 |                                                    |
[2M, 4M)              45 |@                                                   |
[4M, 8M)             857 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@                   |
[8M, 16M)            812 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@                     |
[16M, 32M)          1325 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
[32M, 64M)            31 |@                                                   |

The patch leads to a significant decrease in latency of a newly woken task.

Tested-by: Mukesh Kumar Chaurasiya <mchauras@linux.ibm.com>
Reviewed-by: Mukesh Kumar Chaurasiya <mchauras@linux.ibm.com>