Re: [PATCH v5] erofs: add per-cpu threads for decompression as an option

[Sandeep Dhavale]

On Mon, Feb 27, 2023 at 9:01 PM Gao Xiang <hsiangkao@xxxxxxxxxxxxxxxxx> wrote:
>
> Hi Sandeep,
>
> On 2023/2/28 12:47, Sandeep Dhavale via Linux-erofs wrote:
> > Hi all,
> > I completed the tests and the results are consistent with
> > our previous observation. We can see that removing WQ_UNBOUND
> > helps but the scheduling latency by using high priority per cpu
> > kthreads is even lower. Below is the table.
> >
> > |---------------------+-------+-------+------+-------|
> > | Table               | avg   | med   | min  | max   |
> > |---------------------+-------+-------+------+-------|
> > | Default erofs       | 19323 | 19758 | 3986 | 35051 |
> > |---------------------+-------+-------+------+-------|
> > | !WQ_UNBOUND         | 11202 | 10798 | 3493 | 19822 |
> > |---------------------+-------+-------+------+-------|
> > | hipri pcpu kthreads | 7182  | 7017  | 2463 | 12300 |
> > |---------------------+-------+-------+------+-------|
>
> May I ask did it test with different setup since the test results
> in the original commit message are:
>
Hi Gao,
Yes I did the test on the different (older) hardware than my original testing
(but the same one Nathan had used) to remove that as a variable.

Thanks,
Sandeep.

> +-------------------------+-----------+----------------+---------+
> |                         | workqueue | kthread_worker |  diff   |
> +-------------------------+-----------+----------------+---------+
> | Average (us)            |     15253 |           2914 | -80.89% |
> | Median (us)             |     14001 |           2912 | -79.20% |
> | Minimum (us)            |      3117 |           1027 | -67.05% |
> | Maximum (us)            |     30170 |           3805 | -87.39% |
> | Standard deviation (us) |      7166 |            359 |         |
> +-------------------------+-----------+----------------+---------+
>
> Otherwise it looks good to me for now, hopefully helpful to Android
> users.
>
> Thanks,
> Gao Xiang
>
> >
> >
> > Thanks,
> > Sandeep.