On 2024/9/10 3:04, Michal Koutný wrote:

Hello.

On Thu, Sep 05, 2024 at 01:41:27PM GMT, Chen Ridong [off-list ref] wrote:

quoted

We optimized the freezer to reduce redundant loops. We add a selftest to
ensure our optimizations cause no harm, and we confirmed that the
performance can be improved.

I think these patches have potential to make the code also cleaner and
better to understand (the numbers from optimization don't impress me
that much at the moment).

I post more comments on the individual patches.

Thanks.

quoted

We tested the following subtree: D, E, F and G each have n children.
         A
        / \
       B   C
     / | \  \
     D E F   G
    /  |  \   \
  1-n 1-n 1-n  0-n

(Why is the last one different?)

Sorry, this should be same.

quoted

Our test is to freeze A B C D E F G, and then unfreeze A B C D E F G.

I would say measuring freezing/unfreezing of only A would give better
idea about the impact of the change.
(Going through all descendants manually would unfairly show greater
improvement.)

I will add this test.
And I only test freezing/unfreezing A, The first patch may not be effective.

quoted

We measured the elapsed time.

	BEFORE(ns)	AFTER(ns)	SAVED(ns)
n=10	142679950	139666014	3,013,936
n=100	199832160	192773032	7,059,128
n=1000	488595100	414901570	73,693,530

How many tasks were there inside cgroups in this benchmark?

I didn't put tasks inside cgroups to test. Because this series didn't 
change how the tasks be froze. That only reduced some redundant loop.

(I assume in practice the freezing time would be dominated by waiting
for tasks' response, so it'd be good to note this beside this result.)

Thanks,
Michal