tag:blogger.com,1999:blog-2963790426029213933.post1289577295211305929..comments2024-02-29T00:33:07.382-08:00Comments on Alfred Chen's Blog: New implementaion of skip list for BFSAlfred Chenhttp://www.blogger.com/profile/03164306846702841944noreply@blogger.comBlogger6125tag:blogger.com,1999:blog-2963790426029213933.post-46044713679660711312016-09-17T09:44:58.242-07:002016-09-17T09:44:58.242-07:00That is correct. I tested v4.7_0472_sl_baseline.
I...That is correct. I tested v4.7_0472_sl_baseline.<br />In the meantime I've also run the tests for v4.7_0472_sl_new.<br />When using acpi-cpufreq+performance, the results are the same as v4.7_0472_sl_baseline, and indeed it seems there is a regression under low workload.<br />However, when using acpi-cpufreq+ondemand, the performance is worse in single-threaded tests and in make -j2.<br /><br />PedroAnonymousnoreply@blogger.comtag:blogger.com,1999:blog-2963790426029213933.post-1219056120965406172016-09-16T07:27:53.658-07:002016-09-16T07:27:53.658-07:00@Pedro
As I can see in your result that you only t...@Pedro<br />As I can see in your result that you only test the 4.7 sl_baseline? I guess that's the baseline version in this post?<br />If it is correct, this baseline version is the start point I started my work on and without my embedded and new implementation of SL. There are just 5 commits add upon 472, all in my previous blog. You should also try the second patch(https://bitbucket.org/alfredchen/linux-gc/downloads/v4.7_0472_sl_new.patch), which include all new changes in this post.<br /><br />And if so, based on your result in "make -j2", there maybe performance regression under low workload in 0497, and likely caused by other code change beside skiplist changes.<br />In my testing, pstate driver shows less performance than acpi driver and even the benchmark of mainline kernel in www.phoronix.com shows similar result. So I put it in low priority, good news is CK has picked it up and trying to fix it. But there is too many code changes in 480~497, and code doesn't settle down yet. So I am not going to pick up other code chagnes except skiplist at this moment.Alfred Chenhttps://www.blogger.com/profile/03164306846702841944noreply@blogger.comtag:blogger.com,1999:blog-2963790426029213933.post-30178757308145610952016-09-16T04:15:14.420-07:002016-09-16T04:15:14.420-07:00Hello, and thank you for your work.
Out of curiosi...Hello, and thank you for your work.<br />Out of curiosity I've tested your implementation of bfs, and also CK's bfs 497. I've put the results here:<br />https://docs.google.com/spreadsheets/d/1ZfXUfcP2fBpQA6LLb-DP6xyDgPdFYZMwJdE0SQ6y3Xg/edit?usp=sharing<br /><br />The throughput is on par with bfs 497 when using the acpi-cpufreq+performance governor, and even slightly better on make -j2. However the cpu frequency signaling part seems broken (as it was on CK's bfs 480) because performance is low when using intel_pstate, but I think you are already aware of that.<br />I'll wait for linux 4.8 and for the changes in bfs to settle before doing more tests.<br /><br />PedroAnonymousnoreply@blogger.comtag:blogger.com,1999:blog-2963790426029213933.post-36335055399182619862016-09-16T01:16:39.359-07:002016-09-16T01:16:39.359-07:00@Manuel
There is huge code change in 480~497 and i...@Manuel<br />There is huge code change in 480~497 and it seems that there is still issues with new bfs. So I decided to pick up the skiplist changes from 480/497 first, b/c it has the most value and less risky IMO. And one change at a time helps to isolate the code change when there is problem, and it also help to observe how skiplist impace the scheduler.<br />As 4.8 will be released soon, most likely there will be 0472+skiplist+vrq at the earlier of 4.8 cycle.Alfred Chenhttps://www.blogger.com/profile/03164306846702841944noreply@blogger.comtag:blogger.com,1999:blog-2963790426029213933.post-34934806415802927342016-09-15T20:32:47.828-07:002016-09-15T20:32:47.828-07:00@Alfred:
Btw., why do you still call it 472, when ...@Alfred:<br />Btw., why do you still call it 472, when you're at 497 code level? Sorry, just kidding a little. You might have your reasons.<br /><br />I'm looking forward to your newest revision of good old VRQ on top of skiplist!<br />Current sl_new patchset still works fine.<br />Please keep up your work!<br /><br />BR Manuel KrauseAnonymousnoreply@blogger.comtag:blogger.com,1999:blog-2963790426029213933.post-45086357989008701362016-09-15T09:56:15.795-07:002016-09-15T09:56:15.795-07:00O.k., good work, "sl_new" is up and runn...O.k., good work, "sl_new" is up and running for some hours now. At the moment I want to only post the proof that it's working well on here for now. I haven't faced any issue or interactivity related regression.<br /><br />BR Manuel KrauseAnonymousnoreply@blogger.com