1REDUCING OS JITTER DUE TO PER-CPU KTHREADS 2 3This document lists per-CPU kthreads in the Linux kernel and presents 4options to control their OS jitter. Note that non-per-CPU kthreads are 5not listed here. To reduce OS jitter from non-per-CPU kthreads, bind 6them to a "housekeeping" CPU dedicated to such work. 7 8 9REFERENCES 10 11o Documentation/IRQ-affinity.txt: Binding interrupts to sets of CPUs. 12 13o Documentation/cgroups: Using cgroups to bind tasks to sets of CPUs. 14 15o man taskset: Using the taskset command to bind tasks to sets 16 of CPUs. 17 18o man sched_setaffinity: Using the sched_setaffinity() system 19 call to bind tasks to sets of CPUs. 20 21o /sys/devices/system/cpu/cpuN/online: Control CPU N's hotplug state, 22 writing "0" to offline and "1" to online. 23 24o In order to locate kernel-generated OS jitter on CPU N: 25 26 cd /sys/kernel/debug/tracing 27 echo 1 > max_graph_depth # Increase the "1" for more detail 28 echo function_graph > current_tracer 29 # run workload 30 cat per_cpu/cpuN/trace 31 32 33KTHREADS 34 35Name: ehca_comp/%u 36Purpose: Periodically process Infiniband-related work. 37To reduce its OS jitter, do any of the following: 381. Don't use eHCA Infiniband hardware, instead choosing hardware 39 that does not require per-CPU kthreads. This will prevent these 40 kthreads from being created in the first place. (This will 41 work for most people, as this hardware, though important, is 42 relatively old and is produced in relatively low unit volumes.) 432. Do all eHCA-Infiniband-related work on other CPUs, including 44 interrupts. 453. Rework the eHCA driver so that its per-CPU kthreads are 46 provisioned only on selected CPUs. 47 48 49Name: irq/%d-%s 50Purpose: Handle threaded interrupts. 51To reduce its OS jitter, do the following: 521. Use irq affinity to force the irq threads to execute on 53 some other CPU. 54 55Name: kcmtpd_ctr_%d 56Purpose: Handle Bluetooth work. 57To reduce its OS jitter, do one of the following: 581. Don't use Bluetooth, in which case these kthreads won't be 59 created in the first place. 602. Use irq affinity to force Bluetooth-related interrupts to 61 occur on some other CPU and furthermore initiate all 62 Bluetooth activity on some other CPU. 63 64Name: ksoftirqd/%u 65Purpose: Execute softirq handlers when threaded or when under heavy load. 66To reduce its OS jitter, each softirq vector must be handled 67separately as follows: 68TIMER_SOFTIRQ: Do all of the following: 691. To the extent possible, keep the CPU out of the kernel when it 70 is non-idle, for example, by avoiding system calls and by forcing 71 both kernel threads and interrupts to execute elsewhere. 722. Build with CONFIG_HOTPLUG_CPU=y. After boot completes, force 73 the CPU offline, then bring it back online. This forces 74 recurring timers to migrate elsewhere. If you are concerned 75 with multiple CPUs, force them all offline before bringing the 76 first one back online. Once you have onlined the CPUs in question, 77 do not offline any other CPUs, because doing so could force the 78 timer back onto one of the CPUs in question. 79NET_TX_SOFTIRQ and NET_RX_SOFTIRQ: Do all of the following: 801. Force networking interrupts onto other CPUs. 812. Initiate any network I/O on other CPUs. 823. Once your application has started, prevent CPU-hotplug operations 83 from being initiated from tasks that might run on the CPU to 84 be de-jittered. (It is OK to force this CPU offline and then 85 bring it back online before you start your application.) 86BLOCK_SOFTIRQ: Do all of the following: 871. Force block-device interrupts onto some other CPU. 882. Initiate any block I/O on other CPUs. 893. Once your application has started, prevent CPU-hotplug operations 90 from being initiated from tasks that might run on the CPU to 91 be de-jittered. (It is OK to force this CPU offline and then 92 bring it back online before you start your application.) 93BLOCK_IOPOLL_SOFTIRQ: Do all of the following: 941. Force block-device interrupts onto some other CPU. 952. Initiate any block I/O and block-I/O polling on other CPUs. 963. Once your application has started, prevent CPU-hotplug operations 97 from being initiated from tasks that might run on the CPU to 98 be de-jittered. (It is OK to force this CPU offline and then 99 bring it back online before you start your application.) 100TASKLET_SOFTIRQ: Do one or more of the following: 1011. Avoid use of drivers that use tasklets. (Such drivers will contain 102 calls to things like tasklet_schedule().) 1032. Convert all drivers that you must use from tasklets to workqueues. 1043. Force interrupts for drivers using tasklets onto other CPUs, 105 and also do I/O involving these drivers on other CPUs. 106SCHED_SOFTIRQ: Do all of the following: 1071. Avoid sending scheduler IPIs to the CPU to be de-jittered, 108 for example, ensure that at most one runnable kthread is present 109 on that CPU. If a thread that expects to run on the de-jittered 110 CPU awakens, the scheduler will send an IPI that can result in 111 a subsequent SCHED_SOFTIRQ. 1122. Build with CONFIG_RCU_NOCB_CPU=y, CONFIG_RCU_NOCB_CPU_ALL=y, 113 CONFIG_NO_HZ_FULL=y, and, in addition, ensure that the CPU 114 to be de-jittered is marked as an adaptive-ticks CPU using the 115 "nohz_full=" boot parameter. This reduces the number of 116 scheduler-clock interrupts that the de-jittered CPU receives, 117 minimizing its chances of being selected to do the load balancing 118 work that runs in SCHED_SOFTIRQ context. 1193. To the extent possible, keep the CPU out of the kernel when it 120 is non-idle, for example, by avoiding system calls and by 121 forcing both kernel threads and interrupts to execute elsewhere. 122 This further reduces the number of scheduler-clock interrupts 123 received by the de-jittered CPU. 124HRTIMER_SOFTIRQ: Do all of the following: 1251. To the extent possible, keep the CPU out of the kernel when it 126 is non-idle. For example, avoid system calls and force both 127 kernel threads and interrupts to execute elsewhere. 1282. Build with CONFIG_HOTPLUG_CPU=y. Once boot completes, force the 129 CPU offline, then bring it back online. This forces recurring 130 timers to migrate elsewhere. If you are concerned with multiple 131 CPUs, force them all offline before bringing the first one 132 back online. Once you have onlined the CPUs in question, do not 133 offline any other CPUs, because doing so could force the timer 134 back onto one of the CPUs in question. 135RCU_SOFTIRQ: Do at least one of the following: 1361. Offload callbacks and keep the CPU in either dyntick-idle or 137 adaptive-ticks state by doing all of the following: 138 a. Build with CONFIG_RCU_NOCB_CPU=y, CONFIG_RCU_NOCB_CPU_ALL=y, 139 CONFIG_NO_HZ_FULL=y, and, in addition ensure that the CPU 140 to be de-jittered is marked as an adaptive-ticks CPU using 141 the "nohz_full=" boot parameter. Bind the rcuo kthreads 142 to housekeeping CPUs, which can tolerate OS jitter. 143 b. To the extent possible, keep the CPU out of the kernel 144 when it is non-idle, for example, by avoiding system 145 calls and by forcing both kernel threads and interrupts 146 to execute elsewhere. 1472. Enable RCU to do its processing remotely via dyntick-idle by 148 doing all of the following: 149 a. Build with CONFIG_NO_HZ=y and CONFIG_RCU_FAST_NO_HZ=y. 150 b. Ensure that the CPU goes idle frequently, allowing other 151 CPUs to detect that it has passed through an RCU quiescent 152 state. If the kernel is built with CONFIG_NO_HZ_FULL=y, 153 userspace execution also allows other CPUs to detect that 154 the CPU in question has passed through a quiescent state. 155 c. To the extent possible, keep the CPU out of the kernel 156 when it is non-idle, for example, by avoiding system 157 calls and by forcing both kernel threads and interrupts 158 to execute elsewhere. 159 160Name: kworker/%u:%d%s (cpu, id, priority) 161Purpose: Execute workqueue requests 162To reduce its OS jitter, do any of the following: 1631. Run your workload at a real-time priority, which will allow 164 preempting the kworker daemons. 1652. A given workqueue can be made visible in the sysfs filesystem 166 by passing the WQ_SYSFS to that workqueue's alloc_workqueue(). 167 Such a workqueue can be confined to a given subset of the 168 CPUs using the /sys/devices/virtual/workqueue/*/cpumask sysfs 169 files. The set of WQ_SYSFS workqueues can be displayed using 170 "ls sys/devices/virtual/workqueue". That said, the workqueues 171 maintainer would like to caution people against indiscriminately 172 sprinkling WQ_SYSFS across all the workqueues. The reason for 173 caution is that it is easy to add WQ_SYSFS, but because sysfs is 174 part of the formal user/kernel API, it can be nearly impossible 175 to remove it, even if its addition was a mistake. 1763. Do any of the following needed to avoid jitter that your 177 application cannot tolerate: 178 a. Build your kernel with CONFIG_SLUB=y rather than 179 CONFIG_SLAB=y, thus avoiding the slab allocator's periodic 180 use of each CPU's workqueues to run its cache_reap() 181 function. 182 b. Avoid using oprofile, thus avoiding OS jitter from 183 wq_sync_buffer(). 184 c. Limit your CPU frequency so that a CPU-frequency 185 governor is not required, possibly enlisting the aid of 186 special heatsinks or other cooling technologies. If done 187 correctly, and if you CPU architecture permits, you should 188 be able to build your kernel with CONFIG_CPU_FREQ=n to 189 avoid the CPU-frequency governor periodically running 190 on each CPU, including cs_dbs_timer() and od_dbs_timer(). 191 WARNING: Please check your CPU specifications to 192 make sure that this is safe on your particular system. 193 d. As of v3.18, Christoph Lameter's on-demand vmstat workers 194 commit prevents OS jitter due to vmstat_update() on 195 CONFIG_SMP=y systems. Before v3.18, is not possible 196 to entirely get rid of the OS jitter, but you can 197 decrease its frequency by writing a large value to 198 /proc/sys/vm/stat_interval. The default value is HZ, 199 for an interval of one second. Of course, larger values 200 will make your virtual-memory statistics update more 201 slowly. Of course, you can also run your workload at 202 a real-time priority, thus preempting vmstat_update(), 203 but if your workload is CPU-bound, this is a bad idea. 204 However, there is an RFC patch from Christoph Lameter 205 (based on an earlier one from Gilad Ben-Yossef) that 206 reduces or even eliminates vmstat overhead for some 207 workloads at https://lkml.org/lkml/2013/9/4/379. 208 e. Boot with "elevator=noop" to avoid workqueue use by 209 the block layer. 210 f. If running on high-end powerpc servers, build with 211 CONFIG_PPC_RTAS_DAEMON=n. This prevents the RTAS 212 daemon from running on each CPU every second or so. 213 (This will require editing Kconfig files and will defeat 214 this platform's RAS functionality.) This avoids jitter 215 due to the rtas_event_scan() function. 216 WARNING: Please check your CPU specifications to 217 make sure that this is safe on your particular system. 218 g. If running on Cell Processor, build your kernel with 219 CBE_CPUFREQ_SPU_GOVERNOR=n to avoid OS jitter from 220 spu_gov_work(). 221 WARNING: Please check your CPU specifications to 222 make sure that this is safe on your particular system. 223 h. If running on PowerMAC, build your kernel with 224 CONFIG_PMAC_RACKMETER=n to disable the CPU-meter, 225 avoiding OS jitter from rackmeter_do_timer(). 226 227Name: rcuc/%u 228Purpose: Execute RCU callbacks in CONFIG_RCU_BOOST=y kernels. 229To reduce its OS jitter, do at least one of the following: 2301. Build the kernel with CONFIG_PREEMPT=n. This prevents these 231 kthreads from being created in the first place, and also obviates 232 the need for RCU priority boosting. This approach is feasible 233 for workloads that do not require high degrees of responsiveness. 2342. Build the kernel with CONFIG_RCU_BOOST=n. This prevents these 235 kthreads from being created in the first place. This approach 236 is feasible only if your workload never requires RCU priority 237 boosting, for example, if you ensure frequent idle time on all 238 CPUs that might execute within the kernel. 2393. Build with CONFIG_RCU_NOCB_CPU=y and CONFIG_RCU_NOCB_CPU_ALL=y, 240 which offloads all RCU callbacks to kthreads that can be moved 241 off of CPUs susceptible to OS jitter. This approach prevents the 242 rcuc/%u kthreads from having any work to do, so that they are 243 never awakened. 2444. Ensure that the CPU never enters the kernel, and, in particular, 245 avoid initiating any CPU hotplug operations on this CPU. This is 246 another way of preventing any callbacks from being queued on the 247 CPU, again preventing the rcuc/%u kthreads from having any work 248 to do. 249 250Name: rcuob/%d, rcuop/%d, and rcuos/%d 251Purpose: Offload RCU callbacks from the corresponding CPU. 252To reduce its OS jitter, do at least one of the following: 2531. Use affinity, cgroups, or other mechanism to force these kthreads 254 to execute on some other CPU. 2552. Build with CONFIG_RCU_NOCB_CPU=n, which will prevent these 256 kthreads from being created in the first place. However, please 257 note that this will not eliminate OS jitter, but will instead 258 shift it to RCU_SOFTIRQ. 259 260Name: watchdog/%u 261Purpose: Detect software lockups on each CPU. 262To reduce its OS jitter, do at least one of the following: 2631. Build with CONFIG_LOCKUP_DETECTOR=n, which will prevent these 264 kthreads from being created in the first place. 2652. Boot with "nosoftlockup=0", which will also prevent these kthreads 266 from being created. Other related watchdog and softlockup boot 267 parameters may be found in Documentation/kernel-parameters.txt 268 and Documentation/watchdog/watchdog-parameters.txt. 2693. Echo a zero to /proc/sys/kernel/watchdog to disable the 270 watchdog timer. 2714. Echo a large number of /proc/sys/kernel/watchdog_thresh in 272 order to reduce the frequency of OS jitter due to the watchdog 273 timer down to a level that is acceptable for your workload. 274