1========================================== 2Reducing OS jitter due to per-cpu kthreads 3========================================== 4 5This document lists per-CPU kthreads in the Linux kernel and presents 6options to control their OS jitter. Note that non-per-CPU kthreads are 7not listed here. To reduce OS jitter from non-per-CPU kthreads, bind 8them to a "housekeeping" CPU dedicated to such work. 9 10References 11========== 12 13- Documentation/core-api/irq/irq-affinity.rst: Binding interrupts to sets of CPUs. 14 15- Documentation/admin-guide/cgroup-v1: Using cgroups to bind tasks to sets of CPUs. 16 17- man taskset: Using the taskset command to bind tasks to sets 18 of CPUs. 19 20- man sched_setaffinity: Using the sched_setaffinity() system 21 call to bind tasks to sets of CPUs. 22 23- /sys/devices/system/cpu/cpuN/online: Control CPU N's hotplug state, 24 writing "0" to offline and "1" to online. 25 26- In order to locate kernel-generated OS jitter on CPU N: 27 28 cd /sys/kernel/tracing 29 echo 1 > max_graph_depth # Increase the "1" for more detail 30 echo function_graph > current_tracer 31 # run workload 32 cat per_cpu/cpuN/trace 33 34kthreads 35======== 36 37Name: 38 ehca_comp/%u 39 40Purpose: 41 Periodically process Infiniband-related work. 42 43To reduce its OS jitter, do any of the following: 44 451. Don't use eHCA Infiniband hardware, instead choosing hardware 46 that does not require per-CPU kthreads. This will prevent these 47 kthreads from being created in the first place. (This will 48 work for most people, as this hardware, though important, is 49 relatively old and is produced in relatively low unit volumes.) 502. Do all eHCA-Infiniband-related work on other CPUs, including 51 interrupts. 523. Rework the eHCA driver so that its per-CPU kthreads are 53 provisioned only on selected CPUs. 54 55 56Name: 57 irq/%d-%s 58 59Purpose: 60 Handle threaded interrupts. 61 62To reduce its OS jitter, do the following: 63 641. Use irq affinity to force the irq threads to execute on 65 some other CPU. 66 67Name: 68 kcmtpd_ctr_%d 69 70Purpose: 71 Handle Bluetooth work. 72 73To reduce its OS jitter, do one of the following: 74 751. Don't use Bluetooth, in which case these kthreads won't be 76 created in the first place. 772. Use irq affinity to force Bluetooth-related interrupts to 78 occur on some other CPU and furthermore initiate all 79 Bluetooth activity on some other CPU. 80 81Name: 82 ksoftirqd/%u 83 84Purpose: 85 Execute softirq handlers when threaded or when under heavy load. 86 87To reduce its OS jitter, each softirq vector must be handled 88separately as follows: 89 90TIMER_SOFTIRQ 91------------- 92 93Do all of the following: 94 951. To the extent possible, keep the CPU out of the kernel when it 96 is non-idle, for example, by avoiding system calls and by forcing 97 both kernel threads and interrupts to execute elsewhere. 982. Build with CONFIG_HOTPLUG_CPU=y. After boot completes, force 99 the CPU offline, then bring it back online. This forces 100 recurring timers to migrate elsewhere. If you are concerned 101 with multiple CPUs, force them all offline before bringing the 102 first one back online. Once you have onlined the CPUs in question, 103 do not offline any other CPUs, because doing so could force the 104 timer back onto one of the CPUs in question. 105 106NET_TX_SOFTIRQ and NET_RX_SOFTIRQ 107--------------------------------- 108 109Do all of the following: 110 1111. Force networking interrupts onto other CPUs. 1122. Initiate any network I/O on other CPUs. 1133. Once your application has started, prevent CPU-hotplug operations 114 from being initiated from tasks that might run on the CPU to 115 be de-jittered. (It is OK to force this CPU offline and then 116 bring it back online before you start your application.) 117 118BLOCK_SOFTIRQ 119------------- 120 121Do all of the following: 122 1231. Force block-device interrupts onto some other CPU. 1242. Initiate any block I/O on other CPUs. 1253. Once your application has started, prevent CPU-hotplug operations 126 from being initiated from tasks that might run on the CPU to 127 be de-jittered. (It is OK to force this CPU offline and then 128 bring it back online before you start your application.) 129 130IRQ_POLL_SOFTIRQ 131---------------- 132 133Do all of the following: 134 1351. Force block-device interrupts onto some other CPU. 1362. Initiate any block I/O and block-I/O polling on other CPUs. 1373. Once your application has started, prevent CPU-hotplug operations 138 from being initiated from tasks that might run on the CPU to 139 be de-jittered. (It is OK to force this CPU offline and then 140 bring it back online before you start your application.) 141 142TASKLET_SOFTIRQ 143--------------- 144 145Do one or more of the following: 146 1471. Avoid use of drivers that use tasklets. (Such drivers will contain 148 calls to things like tasklet_schedule().) 1492. Convert all drivers that you must use from tasklets to workqueues. 1503. Force interrupts for drivers using tasklets onto other CPUs, 151 and also do I/O involving these drivers on other CPUs. 152 153SCHED_SOFTIRQ 154------------- 155 156Do all of the following: 157 1581. Avoid sending scheduler IPIs to the CPU to be de-jittered, 159 for example, ensure that at most one runnable kthread is present 160 on that CPU. If a thread that expects to run on the de-jittered 161 CPU awakens, the scheduler will send an IPI that can result in 162 a subsequent SCHED_SOFTIRQ. 1632. CONFIG_NO_HZ_FULL=y and ensure that the CPU to be de-jittered 164 is marked as an adaptive-ticks CPU using the "nohz_full=" 165 boot parameter. This reduces the number of scheduler-clock 166 interrupts that the de-jittered CPU receives, minimizing its 167 chances of being selected to do the load balancing work that 168 runs in SCHED_SOFTIRQ context. 1693. To the extent possible, keep the CPU out of the kernel when it 170 is non-idle, for example, by avoiding system calls and by 171 forcing both kernel threads and interrupts to execute elsewhere. 172 This further reduces the number of scheduler-clock interrupts 173 received by the de-jittered CPU. 174 175HRTIMER_SOFTIRQ 176--------------- 177 178Do all of the following: 179 1801. To the extent possible, keep the CPU out of the kernel when it 181 is non-idle. For example, avoid system calls and force both 182 kernel threads and interrupts to execute elsewhere. 1832. Build with CONFIG_HOTPLUG_CPU=y. Once boot completes, force the 184 CPU offline, then bring it back online. This forces recurring 185 timers to migrate elsewhere. If you are concerned with multiple 186 CPUs, force them all offline before bringing the first one 187 back online. Once you have onlined the CPUs in question, do not 188 offline any other CPUs, because doing so could force the timer 189 back onto one of the CPUs in question. 190 191RCU_SOFTIRQ 192----------- 193 194Do at least one of the following: 195 1961. Offload callbacks and keep the CPU in either dyntick-idle or 197 adaptive-ticks state by doing all of the following: 198 199 a. CONFIG_NO_HZ_FULL=y and ensure that the CPU to be 200 de-jittered is marked as an adaptive-ticks CPU using the 201 "nohz_full=" boot parameter. Bind the rcuo kthreads to 202 housekeeping CPUs, which can tolerate OS jitter. 203 b. To the extent possible, keep the CPU out of the kernel 204 when it is non-idle, for example, by avoiding system 205 calls and by forcing both kernel threads and interrupts 206 to execute elsewhere. 207 2082. Enable RCU to do its processing remotely via dyntick-idle by 209 doing all of the following: 210 211 a. Build with CONFIG_NO_HZ=y. 212 b. Ensure that the CPU goes idle frequently, allowing other 213 CPUs to detect that it has passed through an RCU quiescent 214 state. If the kernel is built with CONFIG_NO_HZ_FULL=y, 215 userspace execution also allows other CPUs to detect that 216 the CPU in question has passed through a quiescent state. 217 c. To the extent possible, keep the CPU out of the kernel 218 when it is non-idle, for example, by avoiding system 219 calls and by forcing both kernel threads and interrupts 220 to execute elsewhere. 221 222Name: 223 kworker/%u:%d%s (cpu, id, priority) 224 225Purpose: 226 Execute workqueue requests 227 228To reduce its OS jitter, do any of the following: 229 2301. Run your workload at a real-time priority, which will allow 231 preempting the kworker daemons. 2322. A given workqueue can be made visible in the sysfs filesystem 233 by passing the WQ_SYSFS to that workqueue's alloc_workqueue(). 234 Such a workqueue can be confined to a given subset of the 235 CPUs using the ``/sys/devices/virtual/workqueue/*/cpumask`` sysfs 236 files. The set of WQ_SYSFS workqueues can be displayed using 237 "ls /sys/devices/virtual/workqueue". That said, the workqueues 238 maintainer would like to caution people against indiscriminately 239 sprinkling WQ_SYSFS across all the workqueues. The reason for 240 caution is that it is easy to add WQ_SYSFS, but because sysfs is 241 part of the formal user/kernel API, it can be nearly impossible 242 to remove it, even if its addition was a mistake. 2433. Do any of the following needed to avoid jitter that your 244 application cannot tolerate: 245 246 a. Avoid using oprofile, thus avoiding OS jitter from 247 wq_sync_buffer(). 248 b. Limit your CPU frequency so that a CPU-frequency 249 governor is not required, possibly enlisting the aid of 250 special heatsinks or other cooling technologies. If done 251 correctly, and if you CPU architecture permits, you should 252 be able to build your kernel with CONFIG_CPU_FREQ=n to 253 avoid the CPU-frequency governor periodically running 254 on each CPU, including cs_dbs_timer() and od_dbs_timer(). 255 256 WARNING: Please check your CPU specifications to 257 make sure that this is safe on your particular system. 258 c. As of v3.18, Christoph Lameter's on-demand vmstat workers 259 commit prevents OS jitter due to vmstat_update() on 260 CONFIG_SMP=y systems. Before v3.18, is not possible 261 to entirely get rid of the OS jitter, but you can 262 decrease its frequency by writing a large value to 263 /proc/sys/vm/stat_interval. The default value is HZ, 264 for an interval of one second. Of course, larger values 265 will make your virtual-memory statistics update more 266 slowly. Of course, you can also run your workload at 267 a real-time priority, thus preempting vmstat_update(), 268 but if your workload is CPU-bound, this is a bad idea. 269 However, there is an RFC patch from Christoph Lameter 270 (based on an earlier one from Gilad Ben-Yossef) that 271 reduces or even eliminates vmstat overhead for some 272 workloads at https://lore.kernel.org/r/00000140e9dfd6bd-40db3d4f-c1be-434f-8132-7820f81bb586-000000@email.amazonses.com. 273 d. If running on high-end powerpc servers, build with 274 CONFIG_PPC_RTAS_DAEMON=n. This prevents the RTAS 275 daemon from running on each CPU every second or so. 276 (This will require editing Kconfig files and will defeat 277 this platform's RAS functionality.) This avoids jitter 278 due to the rtas_event_scan() function. 279 WARNING: Please check your CPU specifications to 280 make sure that this is safe on your particular system. 281 e. If running on Cell Processor, build your kernel with 282 CBE_CPUFREQ_SPU_GOVERNOR=n to avoid OS jitter from 283 spu_gov_work(). 284 WARNING: Please check your CPU specifications to 285 make sure that this is safe on your particular system. 286 f. If running on PowerMAC, build your kernel with 287 CONFIG_PMAC_RACKMETER=n to disable the CPU-meter, 288 avoiding OS jitter from rackmeter_do_timer(). 289 290Name: 291 rcuc/%u 292 293Purpose: 294 Execute RCU callbacks in CONFIG_RCU_BOOST=y kernels. 295 296To reduce its OS jitter, do at least one of the following: 297 2981. Build the kernel with CONFIG_PREEMPT=n. This prevents these 299 kthreads from being created in the first place, and also obviates 300 the need for RCU priority boosting. This approach is feasible 301 for workloads that do not require high degrees of responsiveness. 3022. Build the kernel with CONFIG_RCU_BOOST=n. This prevents these 303 kthreads from being created in the first place. This approach 304 is feasible only if your workload never requires RCU priority 305 boosting, for example, if you ensure frequent idle time on all 306 CPUs that might execute within the kernel. 3073. Build with CONFIG_RCU_NOCB_CPU=y and boot with the rcu_nocbs= 308 boot parameter offloading RCU callbacks from all CPUs susceptible 309 to OS jitter. This approach prevents the rcuc/%u kthreads from 310 having any work to do, so that they are never awakened. 3114. Ensure that the CPU never enters the kernel, and, in particular, 312 avoid initiating any CPU hotplug operations on this CPU. This is 313 another way of preventing any callbacks from being queued on the 314 CPU, again preventing the rcuc/%u kthreads from having any work 315 to do. 316 317Name: 318 rcuop/%d, rcuos/%d, and rcuog/%d 319 320Purpose: 321 Offload RCU callbacks from the corresponding CPU. 322 323To reduce its OS jitter, do at least one of the following: 324 3251. Use affinity, cgroups, or other mechanism to force these kthreads 326 to execute on some other CPU. 3272. Build with CONFIG_RCU_NOCB_CPU=n, which will prevent these 328 kthreads from being created in the first place. However, please 329 note that this will not eliminate OS jitter, but will instead 330 shift it to RCU_SOFTIRQ. 331