I wanted to share my current status after working on the schedutil iowait boost issue for a while now. This is what I consider the best solution, happy for anyone to share thoughts and test results (it's simply impossible to cover them all). I'm hoping to remove some (bad) heuristics that have been in the kernel for a long time and are seemingly impossible to evolve. Since the introduction of these heuristics IO workloads have changed and those heuristics can be removed while only really affecting synthetic benchmarks. A task can set the in_iowait flag before calling schedule() to indicate that it is waiting for IO to complete (traditionally this was block IO only). Initially the reason for introducing the flag was almost cosmetic: To distinguish a CPU idle system that has no tasks to run from a CPU idle system that has tasks waiting for IO in order to continue running. I have outlined the issues with the current iowait behaviour and grouped them into: A) iowait as a performance metric, B) for cpuidle specifically, C) for cpufreq specifically. ====== iowait for performance ====== Issues with using iowait as a performance metric: 1. The definition of iowait altogether is vague at best, it is sprinkled across kernel code. Traditionally is for block io only but grepping the tree for shows many counterexamples. 2. Not every IO task is performance-critical (consider writebacks). ====== cpuidle ====== In the cpuidle governor menu, the number of iowaiting tasks is used to determine an interactivity requirement, the more tasks the CPU has with with iowait set the shallower the state selected will be. Relying on iowait for cpuidle is problematic for the following reasons: 1. There is no guarantee for an iowaiting task to be woken up on the same CPU. 2. The task being in iowait says nothing about the idle duration, we could be selecting shallower states for a long time. 3. The task being in iowait doesn't always imply a performance hit with increased latency. 4. If there is such a performance hit, the number of iowaiting tasks doesn't directly correlate. The cpuidle governor teo shows that the reliance on using iowait isn't critical to performance. So this series replaces menu with teo. ====== cpufreq ====== cpufreq governors had iowait-based behavior since a long time, the initial git release in 2.6.12 with the ondemand governor already had iowait_time essentially counting as busy time and thus increasing frequency. iowait boost in schedutil was introduced by commit ("21ca6d2c52f8 cpufreq: schedutil: Add iowait boosting") with it more or less following intel_pstate's approach. Both intel_pstate and schedutil use the enqueueing of a (FAIR) task on the rq to determine the boost. The more often a task that has iowait set is enqueued the higher the applied boost. The underlying problem is that while a task might only utilize the CPU to a small degree (due to the task mostly blocking for IO), the time in-between enqueue and dequeue might be critical to performance (consider that a storage device has no outstanding requests in the meantime and therefore is idling). The low utilization will lead to a low frequency being selected and thus the task's throughput being sub-optimal. We call this the io utilization problem. Relying on iowait wakeups for cpufreq is problematic for the following reasons: 1. Specifically for asynchronous or multi-threaded IO the assumption that we are missing out on block device throughput because it is idling isn't true. We are often boosting (to an inefficient) frequency just to have another request in a queue. 2. It assumes a task that has waited on IO will send io requests in the future. There are additional schedutil-specific reasons listed in the patch. ====== Per-task iowait boosting ====== There is a general consensus that schedutil iowait boosting needed to at least evolve. Attempts of this have been made, mostly centered around associating the boost to the task that is frequently being enqueued. Unfortunately there are problems to this in addition to the ones mentioned about iowait generally: 1. If the time a task is runnable (so between enqueue and dequeue) is critical to throughput, so is the interrupt handling and the block layer completion. For a simple fio example that goes to userspace and back everytime even for a block device driver like nvme-pci that is highly optimised, more than half of the time between io requests is spent on the CPU without the task being enqueued. If the CPU is only boosted for the time that the task is runnable, there is little gain. For !MCQ devices we cannot assume the non-task part migrates with the task, so migration even to a higher capacity CPU might not be beneficial. (This is getting less important though with widespread MCQ devices, see Bart's comments in [1].) 2. If the task benefits from boosting it will be very short-lived (otherwise it would build up utilization on it's own), much shorter than hardware can switch frequency, thus the boost won't be effective until at least the next enqueue-dequeue cycle. Together with the previous reason this means we have to hold the boosted frequency even if the task is off the rq. 3. The increased complexity of implementation. I don't think a per-task io boost strategy that is practical and improves enough on the per-rq design. See my proposals [1] and the OSPM24 discussion [2] about iowait boosting. ====== Performance impact ====== Overall performance impact will depend to some degree on the workload and the platform. Performance is compared for mainline menu and mainline teo with schedutil with or without iowait boost. ----- Synthetic benchmark ----- The worst case is a simple fio benchmark with minimal time spend in userspace and IOs issued sequentially back-to-back. fio --minimal --time_based --name=fiotest --filename=/dev/nvme0n1 --runtime=30 --rw=randread --bs=4k --ioengine=psync --iodepth=1 --direct=1 | cut -f 8 -d ";" The fio task will not exceed utilization for anything but the lowest OPP being requested on an otherwise idle system. On a RK3399 which has high latency incurring sleep states (reported as state1:370us state2:900us, observed on average more like state1:50us state2:100us) using teo can actually offset the performance loss incurred by not using schedutil iowait boosting. RK3399: menu iowait boost /dev/nvme0n1 [2584, 2622, 2629] menu no iowait boost /dev/nvme0n1 [2068, 2097, 2647] teo iowait boost /dev/nvme0n1 [4476, 4478, 4492] teo no iowait boost /dev/nvme0n1 [3728, 3733, 3737] Menu despite the performance multiplier of iowait will often select a deeper state even if the current workload makes these a worse choice for both throughput and energy-efficiency. On a M1 which has just state1 during which the hardware will power-gate WFI which leads to exit latencies of 10us to 20us there is no gain due to using teo, so a regression is visible altogether: M1 Pro: menu iowait boost /dev/nvme0n1 [12780, 12788, 12790] menu no iowait boost /dev/nvme0n1 [8354, 8359, 8389] teo iowait boost /dev/nvme0n1 [12680, 12708, 12728] teo no iowait boost /dev/nvme0n1 [7981, 8005, 8093] One can see the diminishing effect of iowait boosting, even for the synchronous single-threaded case if the workload builds up utilization on it's own by using the --thinkcycles fio parameter. ----- Benchmarks ----- Fortunately most real-world io workloads don't suffer from the io utilization problem. To illustrate here are a few, all done on the M1 Pro again. All are run with default mmtests configs except for the filebench ones where I have set FILEBENCH_MIN_THREADS=8 and FILEBENCH_MAX_THREADS=16. M1 Pro has 8 CPUs. The test are run with mainline menu and mainline teo and the sugov iowait boost knob on or off: menu + iowait_boost, menu - iowait_boost, teo + iowait_boost, teo - iowait_boost, menu + iowait_boost The last column is to get a feel on the error margin of the test. config-db-pgbench-timed-rw-large pgbench Transactions Hmean 1 726.68 ( 0.00%) 713.79 ( -1.77%) 764.83 ( 5.25%) 686.74 ( -5.50%) 715.22 ( -1.58%) Hmean 4 2213.45 ( 0.00%) 2302.69 ( 4.03%) 2155.90 ( -2.60%) 2716.60 * 22.73%* 2099.55 ( -5.15%) Hmean 7 3048.96 ( 0.00%) 2530.59 * -17.00%* 2761.00 ( -9.44%) 2601.78 * -14.67%* 3967.59 * 30.13%* Hmean 12 3223.91 ( 0.00%) 3166.93 ( -1.77%) 3503.41 ( 8.67%) 3122.70 ( -3.14%) 3228.43 ( 0.14%) Hmean 16 3510.04 ( 0.00%) 3405.16 ( -2.99%) 3462.12 ( -1.37%) 3584.56 ( 2.12%) 3476.20 ( -0.96%) The multithreaded tests look good, the upper 3 aren't stable on this platform IMV. I have seen pgbench run fine even for 1 and 4 with teo, the original io_uring regression seemed to be from using menu. config-io-blogbench blogbench Hmean WriteScore 12060.71 ( 0.00%) 12179.88 ( 0.99%) 12155.41 ( 0.79%) 11936.12 ( -1.03%) 12159.15 ( 0.82%) Hmean ReadScore 206741.40 ( 0.00%) 205900.19 ( -0.41%) 204815.61 ( -0.93%) 220167.94 ( 6.49%) 207404.34 ( 0.32%) config-io-bonniepp-dir-async bonnie Throughput Hmean SeqCreate ops 742.63 ( 0.00%) 756.90 ( 1.92%) 784.49 ( 5.64%) 734.96 ( -1.03%) 762.66 ( 2.70%) Hmean SeqCreate read 637.83 ( 0.00%) 539.57 * -15.40%* 645.19 ( 1.15%) 533.63 * -16.34%* 642.83 ( 0.78%) Hmean SeqCreate del 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) Hmean RandCreate ops 1456.37 ( 0.00%) 1418.19 ( -2.62%) 1465.71 ( 0.64%) 1484.20 ( 1.91%) 1552.98 ( 6.63%) Hmean RandCreate read 595.00 ( 0.00%) 581.01 ( -2.35%) 563.78 ( -5.25%) 563.58 ( -5.28%) 562.95 ( -5.39%) The SeqCreate read line looks like a ~15% regression wihout iowait boost. config-io-filebench-varmail-medium filebench Hmean varmail-8 27995.82 ( 0.00%) 27463.93 ( -1.90%) 28341.78 ( 1.24%) 27237.85 * -2.71%* 28164.97 ( 0.60%) Hmean varmail-16 42904.57 ( 0.00%) 41633.66 * -2.96%* 41641.88 * -2.94%* 40831.48 * -4.83%* 41756.76 ( -2.68%) config-io-filebench-webproxy-medium filebench Hmean webproxy-8 191686.03 ( 0.00%) 190237.03 ( -0.76%) 192915.82 ( 0.64%) 190706.94 ( -0.51%) 192282.81 ( 0.31%) Hmean webproxy-16 181528.22 ( 0.00%) 185743.13 * 2.32%* 185432.54 * 2.15%* 186597.08 * 2.79%* 183376.19 ( 1.02%) config-io-filebench-webserver-medium filebench Hmean webserver-4 49974.72 ( 0.00%) 44327.69 * -11.30%* 49498.08 ( -0.95%) 44223.22 * -11.51%* 50726.92 * 1.51%* Hmean webserver-6 73232.09 ( 0.00%) 65999.99 * -9.88%* 73396.77 ( 0.22%) 65628.80 * -10.38%* 72996.77 ( -0.32%) Hmean webserver-12 149138.05 ( 0.00%) 148735.35 ( -0.27%) 148853.11 ( -0.19%) 148119.96 * -0.68%* 148458.42 ( -0.46%) If threads < nr_cpus then there's a ~10% regression due to iowait boosting. config-io-fio-io_uring-randread-direct fio Throughput Hmean kb/sec-randread-read 459665.40 ( 0.00%) 460498.37 ( 0.18%) 442021.24 * -3.84%* 449985.77 * -2.11%* 459446.89 ( -0.05%) Indicates a slight teo regression. config-io-fsmark-xfsrepair fsmark Hmean 4-files/sec 203535.32 ( 0.00%) 202487.47 ( -0.51%) 208261.00 ( 2.32%) 211015.56 * 3.68%* 208432.18 * 2.41%* config-workload-ebizzy ebizzy Overall Throughput Hmean Rsec-1 112650.58 ( 0.00%) 117498.40 ( 4.30%) 112201.69 ( -0.40%) 112297.30 ( -0.31%) 111077.71 ( -1.40%) Hmean Rsec-3 413238.27 ( 0.00%) 385083.09 ( -6.81%) 407349.16 ( -1.43%) 418227.30 ( 1.21%) 418465.17 ( 1.26%) Hmean Rsec-5 638089.98 ( 0.00%) 637027.84 ( -0.17%) 637337.67 ( -0.12%) 639066.37 ( 0.15%) 638654.41 ( 0.09%) Hmean Rsec-7 761622.12 ( 0.00%) 761506.36 ( -0.02%) 761725.47 ( 0.01%) 760639.83 ( -0.13%) 761642.36 ( 0.00%) Hmean Rsec-12 783093.40 ( 0.00%) 765566.10 ( -2.24%) 786634.19 ( 0.45%) 792604.59 ( 1.21%) 796267.24 ( 1.68%) Hmean Rsec-18 784677.51 ( 0.00%) 783555.79 ( -0.14%) 785633.74 ( 0.12%) 771189.93 ( -1.72%) 790559.21 ( 0.75%) Hmean Rsec-24 779544.14 ( 0.00%) 781893.47 ( 0.30%) 777973.36 ( -0.20%) 776038.68 ( -0.45%) 775678.94 ( -0.50%) Hmean Rsec-30 757447.59 ( 0.00%) 733989.03 ( -3.10%) 729435.88 ( -3.70%) 746541.43 ( -1.44%) 752617.47 ( -0.64%) Hmean Rsec-32 760912.43 ( 0.00%) 773417.23 ( 1.64%) 735356.71 ( -3.36%) 743558.83 ( -2.28%) 751862.45 ( -1.19%) No obvious trend here, but since it was used a couple times for cpuidle governor development in the past I did include it. Note: Technically the entire series saves some time on the enqueue path, that is not reflected here as it's all measured with the sysfs knob. ====== RFT ====== The series is structured in a way to make bisecting easy in case of regressions. Additionally I provided two test patches in schedutil to cap the boost using a sysfs knob. In case there are regressions I would be interested if setting a relatively low cap mitigates the issue. I can provide a similar knob for intel_state if regressions are reported that bisect to that. Information about the CPU capacities / topology would be useful if it bisects to the cpufreq changes, about the idle states if it bisects to cpuidle respectively. ====== TODO ====== I haven't touched cpufreq_ondemand yet, that would need to be done as well, but wanted to get some comments and/or testresults first. should_io_be_busy() occurrences also still needs to be addressed. [1] v1 per-task io boost https://lore.kernel.org/lkml/20240304201625.100619-1-christian.loehle@xxxxxxx/ v2 per-task io boost https://lore.kernel.org/lkml/20240518113947.2127802-2-christian.loehle@xxxxxxx/ [2] OSPM24 discussion iowait boosting https://www.youtube.com/watch?v=MSQGEsSziZ4 Regards, Christian Christian Loehle (8): cpuidle: menu: Remove iowait influence cpuidle: Prefer teo over menu governor TEST: cpufreq/schedutil: Linear iowait boost step TEST: cpufreq/schedutil: iowait boost cap sysfs cpufreq/schedutil: Remove iowait boost cpufreq: intel_pstate: Remove iowait boost cpufreq: Remove SCHED_CPUFREQ_IOWAIT update io_uring: Do not set iowait before sleeping drivers/cpufreq/intel_pstate.c | 50 +---------- drivers/cpuidle/Kconfig | 5 +- drivers/cpuidle/governors/menu.c | 78 +++-------------- drivers/cpuidle/governors/teo.c | 2 +- include/linux/sched/cpufreq.h | 2 - io_uring/io_uring.c | 17 ---- kernel/sched/cpufreq_schedutil.c | 144 +------------------------------ kernel/sched/fair.c | 8 -- 8 files changed, 18 insertions(+), 288 deletions(-) -- 2.34.1