RDB + Autogroup testing

[hamadmarri]

Could you please test RDB + Autogroup

I attached 2 patch files

cacule-rdb-autogroup.patch is the full patch cacule+rdb+autogroup
rdb-autogroup-testing.patch is only rdb + autogroup <- in case you need to patch on top of your current cacule

cacule-rdb-autogroup.zip
rdb-autogroup-testing.zip

Thank you

[JohnyPeaN]

@hamadmarri
I have tested it remotely on my home computer and have first observation.
Under load (compiling kernel) rdb doesn't use cores up to 100% and lets them partially run on idle, it can be seen in htop and other utils. Frequency scaling seems good, but i still use ondemand here. Might be some accounting problem, because user+sys are lower for the same work. But the job still took longer so the efficiency seems worse:

5.13.3-181-tkg-cacule RDB AUTOGROUP (output of dstat)
--total-cpu-usage-- -dsk/total- -net/total- ------memory-usage----- ---paging-- ---system-- ---procs---
usr sys idl wai stl| read writ| recv send| used free buff cach| in out | int csw |run blk new|
76 7 16 0 0| 0 0 | 506B 820B|12.4G 48.5G 174M 10.6G| 0 0 | 14k 8982 | 25 0 414|
75 8 17 0 0| 0 0 | 374B 2094B|12.2G 48.7G 174M 10.6G| 0 0 | 14k 9651 | 25 0 488|
78 8 14 0 0| 0 0 | 126B 702B|12.1G 48.7G 174M 10.6G| 0 0 | 13k 8243 | 24 0 388|
76 8 17 0 0| 0 0 | 186B 702B|12.0G 48.9G 174M 10.6G| 0 0 | 14k 9504 | 24 0 499|
72 6 22 0 0| 0 0 | 66B 678B|12.8G 48.1G 174M 10.6G| 0 0 | 12k 8184 | 24 0 347|
63 6 31 0 0| 0 0 | 309B 702B|12.3G 48.6G 174M 10.6G| 0 0 | 11k 7219 | 26 0 308|
77 7 16 0 0| 0 0 | 126B 678B|12.4G 48.5G 174M 10.7G| 0 0 | 13k 8344 | 25 0 366|
73 7 20 0 0| 0 0 | 66B 678B|12.4G 48.5G 174M 10.7G| 0 0 | 12k 7405 | 24 0 333|
real 11m50,965s
user 195m20,447s
sys 18m54,226s

5.13.3-181-tkg-cacule without RDB
--total-cpu-usage-- -dsk/total- -net/total- ------memory-usage----- ---paging-- ---system-- ---procs---
usr sys idl wai stl| read writ| recv send| used free buff cach| in out | int csw |run blk new|
91 8 1 0 0| 0 0 |1775B 2259B|10.9G 48.4G 153M 10.9G| 0 0 | 15k 7119 | 24 0 369|
91 8 1 0 0| 0 0 |1649B 3375B|11.1G 48.2G 153M 10.9G| 0 0 | 14k 7104 | 51 0 355|
90 9 2 0 0| 0 0 |2171B 7069B|11.2G 48.0G 153M 10.9G| 0 0 | 15k 8410 | 25 0 485|
91 8 1 0 0| 0 24k|3598B 4114B|11.1G 48.2G 153M 10.9G| 0 0 | 15k 8275 | 25 0 388|
91 8 1 0 0| 0 0 |1529B 2149B|10.9G 48.3G 153M 11.0G| 0 0 | 15k 7816 | 24 0 398|
91 8 1 0 0| 0 0 | 246B 678B|10.9G 48.3G 153M 11.0G| 0 0 | 14k 6958 | 24 0 384|
91 8 1 0 0| 0 0 |1709B 2165B|10.9G 48.4G 153M 11.0G| 0 0 | 16k 9673 | 24 0 373|
91 8 1 0 0| 0 0 |1709B 2141B|11.0G 48.3G 153M 11.0G| 0 0 | 15k 7410 | 24 0 351|
real 10m34,440s
user 215m24,220s
sys 20m54,322s

I will test responsiveness later.

[ptr1337]

Faced the same issue by me at compiling or also at stress-ng.

[hamadmarri]

@JohnyPeaN @ptr1337

Thank you for testing. I am also having this issue, and I can't yet find the reason why the poor balance. It could be because of:

• head_IS is not updated correctly or propagated through parents group
• rdb_task function is not properly traversing the right children of the group

I can't think of other reasons. Please let me know if you have any proposed idea.

Thank you so much

[raykzhao]

Hi @hamadmarri,

I have noticed that the select_task_rq_fair has been reverted back to the old ways. Just in case someone would like to test the rdb-autogroup with the latest CacULE, I have made the synced patch (apply on top of the latest CacULE).
rdb-autogroup-testing_synced.zip

[hamadmarri]

Hi @hamadmarri,

I have noticed that the select_task_rq_fair has been reverted back to the old ways. Just in case someone would like to test the rdb-autogroup with the latest CacULE, I have made the synced patch (apply on top of the latest CacULE).
rdb-autogroup-testing_synced.zip

Hi @raykzhao

Yes, since the CFS stats and accounting are used, it is better to use normal cacule/cfs way for select_task_rq. Does it have any effects on autogroup patch?

Thank you for providing the patch.

[raykzhao]

Hi @hamadmarri,

On my laptop with this patch the system feels slightly smoother than the previous rdb+autogroup. However, I'm not sure if this patch fixes the issues pointed out by @JohnyPeaN and @ptr1337.

[hamadmarri]

I got some good rdb+autogroup version. It is finally balancing well as plain rdb. The performance also very good and some times better than the old plain rdb. Please test.

Git tree:
https://github.com/hamadmarri/linux/tree/cacule-5.13-rdb-autogroup-testing

All in one (cacule + rdb + autogroup):
https://github.com/hamadmarri/cacule-cpu-scheduler/blob/master/patches/CacULE/experimental/cacule-5.13-rdb-autogroup.patch

Thank you