On Sun, Sep 03, 2023 at 04:06:15PM +0200, Marcus Seyfarth wrote: > > > I should also probably add my RCU config settings here: > > > > > > # RCU Subsystem > > > # > > > CONFIG_TREE_RCU=y > > > CONFIG_PREEMPT_RCU=y > > > CONFIG_RCU_EXPERT=y > > > CONFIG_TREE_SRCU=y > > > CONFIG_TASKS_RCU_GENERIC=y > > > # CONFIG_FORCE_TASKS_RCU is not set > > > CONFIG_TASKS_RCU=y > > > # CONFIG_FORCE_TASKS_RUDE_RCU is not set > > > CONFIG_TASKS_RUDE_RCU=y > > > # CONFIG_FORCE_TASKS_TRACE_RCU is not set > > > CONFIG_TASKS_TRACE_RCU=y > > > CONFIG_RCU_STALL_COMMON=y > > > CONFIG_RCU_NEED_SEGCBLIST=y > > > CONFIG_RCU_FANOUT=32 > > > CONFIG_RCU_FANOUT_LEAF=32 > > > > This CONFIG_RCU_FANOUT_LEAF=32 could result in lock contention, but much > > depends on your workload. > > I followed advise of ChatGPT 3.5 on that setting (while it also warned me > about the potential lock contention, but I haven't observed performance > problems in my benchmarking): > > ChatGPT: "Since the previous recommendation for the CONFIG_RCU_FANOUT > option for the Intel Xeon E5-2696 v3 processor was to start with a value of > 32, it would be appropriate to set CONFIG_RCU_FANOUT_LEAF to the same value > of 32. This aligns the fanout value at both the inner levels and the leaf > level of the RCU hierarchy. Keeping CONFIG_RCU_FANOUT_LEAF consistent with > CONFIG_RCU_FANOUT helps maintain a balanced distribution of resources and > avoids potential performance bottlenecks." Me, I would not trust ChatGPT for this sort of thing, but it is your life, so feel free to do what you want. But if you for whatever reason trust ChatGPT, why don't you ask it for the root cause of the slowdown and the the warning? Why bother asking me? > > CONFIG_RCU_BOOST=y > > > CONFIG_RCU_BOOST_DELAY=2 > > > > Boosting after only two milliseconds is supported, but aggressive. > > That was also advised for gaming workloads by ChatGPT 3.5 but not as > strongly: > > ChatGPT: "For gaming performance, the impact of CONFIG_RCU_BOOST_DELAY is > usually minimal, and the default value should be sufficient for most > scenarios. The default value is typically set to 4 milliseconds (ms) in > many kernel configurations. If you wish to experiment with different values > to optimize gaming performance, you can try reducing the > CONFIG_RCU_BOOST_DELAY value slightly. For example, you can experiment with > values like 2 ms or even 1 ms to potentially improve responsiveness for > real-time tasks, including the game itself." Much of ChatGPT's response is counter-factual, as in ChatGPT just made a fool of itself here, as it has done quite often in the past. Please do everyone (and especially yourself) a big favor and stop trusting its statistical word-salad mashups regarding RCU. Maybe the day will come when AI understands RCU, but today is not that day. > > > # CONFIG_RCU_EXP_KTHREAD is not set > > > CONFIG_RCU_NOCB_CPU=y > > > CONFIG_RCU_NOCB_CPU_DEFAULT_ALL=y > > > CONFIG_RCU_NOCB_CPU_CB_BOOST=y > > > # CONFIG_TASKS_TRACE_RCU_READ_MB is not set > > > CONFIG_RCU_LAZY=y > > > > As noted earlier, please try CONFIG_RCU_LAZY=y. > > As CONFIG_RCU_LAZY=y is set above, I am confused by this suggestion. Do you > want me to unset the other settings? Apologies. I want you to follow the advice I gave you in my earlier email, the one with Message-ID c054b588-b029-4380-9ec5-4ae50ee37d08@paulmck-laptop. I would give you a URL, but you seem to have dropped the public email lists. I have added them back. > > > # CONFIG_RCU_DOUBLE_CHECK_CB_TIME is not set > > > # end of RCU Subsystem > > > > > > I also added this to the CachyOS config file: > > > > > > # Suggestions from BARD for my Xeon E5-2696V3 > > > kernel.rcu_nocbs = 0 > > > > Well and good. > > > > > kernel.rcu_cpu_stall_count = 18 > > > kernel.rcu_queue_length = 1024 > > > kernel.rcu_interval = 200 > > > > I might be having a bad code-search day, but I don't see any sign of > > any of these in mainline or in -rcu. > > > > Do any of the patches that added these also add a call_rcu() that is > > invoked during the time that you observe the slowdowns? More generally, > > I suggest inspecting these patches carefully. I did not take any of > > them into account when reviewing recent changes (nor should I or any of > > the other RCU maintainers or reviewers be expected to), so it is quite > > possible that a recent change invalidated one of those patches. > > These were not a result of any RCU extra patches. These four entries were > added by me manually to the CachyOS-Settings file > (/etc/sysctl.d/99-cachyos-settings.conf) that controls various knobs on > that distro in this central place. Unless there are additional patches (perhaps generated by ChatGPT?), those settings do nothing. > The only patch CachyOS ships with 6.5 that could be somewhat related is: > https://lore.kernel.org/linux-mm/20230703184928.GB4378@monkey/T/#m36fa1f2a52341d57a8ac39f5bd2d64376d26bfe5 > > >From the discussion, I see that it was problematic. I will delete it in my > next experiments. > > The CachyOS Kernel default RCU settings are available under this link: > https://github.com/CachyOS/linux-cachyos/blob/fa4eda73dd00e29fad3c98d49a8843d813b1c1fe/linux-cachyos/config#L163 > [which also reproduced the problems with the default distro Kernel 6.5.0-2]. > > Thanks a lot for your help! You are welcome, but at this point, I am afraid that you are on your own. You have the CONFIG_RCU_LAZY=n suggestion and the bisection suggestion. I wish you the best of everything in your quest. If this problem occurs in mainline, someone will reproduce it reasonably soon against a clean mainline or -stable release. Thanx, Paul
