1 ==========================================
2 Reducing OS jitter due to per-cpu kthreads
3 ==========================================
4
5 This document lists per-CPU kthreads in the Li
6 options to control their OS jitter. Note that
7 not listed here. To reduce OS jitter from non
8 them to a "housekeeping" CPU dedicated to such
9
10 References
11 ==========
12
13 - Documentation/core-api/irq/irq-affinit
14
15 - Documentation/admin-guide/cgroup-v1:
16
17 - man taskset: Using the taskset comman
18 of CPUs.
19
20 - man sched_setaffinity: Using the sche
21 call to bind tasks to sets of CPUs.
22
23 - /sys/devices/system/cpu/cpuN/online:
24 writing "0" to offline and "1" to onli
25
26 - In order to locate kernel-generated OS
27
28 cd /sys/kernel/tracing
29 echo 1 > max_graph_depth # Inc
30 echo function_graph > current_
31 # run workload
32 cat per_cpu/cpuN/trace
33
34 kthreads
35 ========
36
37 Name:
38 ehca_comp/%u
39
40 Purpose:
41 Periodically process Infiniband-related work
42
43 To reduce its OS jitter, do any of the followi
44
45 1. Don't use eHCA Infiniband hardware, in
46 that does not require per-CPU kthreads
47 kthreads from being created in the fir
48 work for most people, as this hardware
49 relatively old and is produced in rela
50 2. Do all eHCA-Infiniband-related work on
51 interrupts.
52 3. Rework the eHCA driver so that its per
53 provisioned only on selected CPUs.
54
55
56 Name:
57 irq/%d-%s
58
59 Purpose:
60 Handle threaded interrupts.
61
62 To reduce its OS jitter, do the following:
63
64 1. Use irq affinity to force the irq thre
65 some other CPU.
66
67 Name:
68 kcmtpd_ctr_%d
69
70 Purpose:
71 Handle Bluetooth work.
72
73 To reduce its OS jitter, do one of the followi
74
75 1. Don't use Bluetooth, in which case the
76 created in the first place.
77 2. Use irq affinity to force Bluetooth-re
78 occur on some other CPU and furthermor
79 Bluetooth activity on some other CPU.
80
81 Name:
82 ksoftirqd/%u
83
84 Purpose:
85 Execute softirq handlers when threaded or wh
86
87 To reduce its OS jitter, each softirq vector m
88 separately as follows:
89
90 TIMER_SOFTIRQ
91 -------------
92
93 Do all of the following:
94
95 1. To the extent possible, keep the CPU o
96 is non-idle, for example, by avoiding
97 both kernel threads and interrupts to
98 2. Build with CONFIG_HOTPLUG_CPU=y. Afte
99 the CPU offline, then bring it back on
100 recurring timers to migrate elsewhere.
101 with multiple CPUs, force them all off
102 first one back online. Once you have
103 do not offline any other CPUs, because
104 timer back onto one of the CPUs in que
105
106 NET_TX_SOFTIRQ and NET_RX_SOFTIRQ
107 ---------------------------------
108
109 Do all of the following:
110
111 1. Force networking interrupts onto other
112 2. Initiate any network I/O on other CPUs
113 3. Once your application has started, pre
114 from being initiated from tasks that m
115 be de-jittered. (It is OK to force th
116 bring it back online before you start
117
118 BLOCK_SOFTIRQ
119 -------------
120
121 Do all of the following:
122
123 1. Force block-device interrupts onto som
124 2. Initiate any block I/O on other CPUs.
125 3. Once your application has started, pre
126 from being initiated from tasks that m
127 be de-jittered. (It is OK to force th
128 bring it back online before you start
129
130 IRQ_POLL_SOFTIRQ
131 ----------------
132
133 Do all of the following:
134
135 1. Force block-device interrupts onto som
136 2. Initiate any block I/O and block-I/O p
137 3. Once your application has started, pre
138 from being initiated from tasks that m
139 be de-jittered. (It is OK to force th
140 bring it back online before you start
141
142 TASKLET_SOFTIRQ
143 ---------------
144
145 Do one or more of the following:
146
147 1. Avoid use of drivers that use tasklets
148 calls to things like tasklet_schedule(
149 2. Convert all drivers that you must use
150 3. Force interrupts for drivers using tas
151 and also do I/O involving these driver
152
153 SCHED_SOFTIRQ
154 -------------
155
156 Do all of the following:
157
158 1. Avoid sending scheduler IPIs to the CP
159 for example, ensure that at most one r
160 on that CPU. If a thread that expects
161 CPU awakens, the scheduler will send a
162 a subsequent SCHED_SOFTIRQ.
163 2. CONFIG_NO_HZ_FULL=y and ensure that th
164 is marked as an adaptive-ticks CPU usi
165 boot parameter. This reduces the numb
166 interrupts that the de-jittered CPU re
167 chances of being selected to do the lo
168 runs in SCHED_SOFTIRQ context.
169 3. To the extent possible, keep the CPU o
170 is non-idle, for example, by avoiding
171 forcing both kernel threads and interr
172 This further reduces the number of sch
173 received by the de-jittered CPU.
174
175 HRTIMER_SOFTIRQ
176 ---------------
177
178 Do all of the following:
179
180 1. To the extent possible, keep the CPU o
181 is non-idle. For example, avoid syste
182 kernel threads and interrupts to execu
183 2. Build with CONFIG_HOTPLUG_CPU=y. Once
184 CPU offline, then bring it back online
185 timers to migrate elsewhere. If you a
186 CPUs, force them all offline before br
187 back online. Once you have onlined th
188 offline any other CPUs, because doing
189 back onto one of the CPUs in question.
190
191 RCU_SOFTIRQ
192 -----------
193
194 Do at least one of the following:
195
196 1. Offload callbacks and keep the CPU in
197 adaptive-ticks state by doing all of t
198
199 a. CONFIG_NO_HZ_FULL=y and ensure
200 de-jittered is marked as an ad
201 "nohz_full=" boot parameter.
202 housekeeping CPUs, which can t
203 b. To the extent possible, keep t
204 when it is non-idle, for examp
205 calls and by forcing both kern
206 to execute elsewhere.
207
208 2. Enable RCU to do its processing remote
209 doing all of the following:
210
211 a. Build with CONFIG_NO_HZ=y.
212 b. Ensure that the CPU goes idle
213 CPUs to detect that it has pas
214 state. If the kernel is built
215 userspace execution also allow
216 the CPU in question has passed
217 c. To the extent possible, keep t
218 when it is non-idle, for examp
219 calls and by forcing both kern
220 to execute elsewhere.
221
222 Name:
223 kworker/%u:%d%s (cpu, id, priority)
224
225 Purpose:
226 Execute workqueue requests
227
228 To reduce its OS jitter, do any of the followi
229
230 1. Run your workload at a real-time prior
231 preempting the kworker daemons.
232 2. A given workqueue can be made visible
233 by passing the WQ_SYSFS to that workqu
234 Such a workqueue can be confined to a
235 CPUs using the ``/sys/devices/virtual/
236 files. The set of WQ_SYSFS workqueues
237 "ls /sys/devices/virtual/workqueue".
238 maintainer would like to caution peopl
239 sprinkling WQ_SYSFS across all the wor
240 caution is that it is easy to add WQ_S
241 part of the formal user/kernel API, it
242 to remove it, even if its addition was
243 3. Do any of the following needed to avoi
244 application cannot tolerate:
245
246 a. Avoid using oprofile, thus avo
247 wq_sync_buffer().
248 b. Limit your CPU frequency so th
249 governor is not required, poss
250 special heatsinks or other coo
251 correctly, and if you CPU arch
252 be able to build your kernel w
253 avoid the CPU-frequency govern
254 on each CPU, including cs_dbs_
255
256 WARNING: Please check your CP
257 make sure that this is safe on
258 c. As of v3.18, Christoph Lameter
259 commit prevents OS jitter due
260 CONFIG_SMP=y systems. Before
261 to entirely get rid of the OS
262 decrease its frequency by writ
263 /proc/sys/vm/stat_interval. T
264 for an interval of one second.
265 will make your virtual-memory
266 slowly. Of course, you can al
267 a real-time priority, thus pre
268 but if your workload is CPU-bo
269 However, there is an RFC patch
270 (based on an earlier one from
271 reduces or even eliminates vms
272 workloads at https://lore.kern
273 d. If running on high-end powerpc
274 CONFIG_PPC_RTAS_DAEMON=n. Thi
275 daemon from running on each CP
276 (This will require editing Kco
277 this platform's RAS functional
278 due to the rtas_event_scan() f
279 WARNING: Please check your CP
280 make sure that this is safe on
281 e. If running on Cell Processor,
282 CBE_CPUFREQ_SPU_GOVERNOR=n to
283 spu_gov_work().
284 WARNING: Please check your CP
285 make sure that this is safe on
286 f. If running on PowerMAC, build
287 CONFIG_PMAC_RACKMETER=n to dis
288 avoiding OS jitter from rackme
289
290 Name:
291 rcuc/%u
292
293 Purpose:
294 Execute RCU callbacks in CONFIG_RCU_BOOST=y
295
296 To reduce its OS jitter, do at least one of th
297
298 1. Build the kernel with CONFIG_PREEMPT=n
299 kthreads from being created in the fir
300 the need for RCU priority boosting. T
301 for workloads that do not require high
302 2. Build the kernel with CONFIG_RCU_BOOST
303 kthreads from being created in the fir
304 is feasible only if your workload neve
305 boosting, for example, if you ensure f
306 CPUs that might execute within the ker
307 3. Build with CONFIG_RCU_NOCB_CPU=y and b
308 boot parameter offloading RCU callback
309 to OS jitter. This approach prevents
310 having any work to do, so that they ar
311 4. Ensure that the CPU never enters the k
312 avoid initiating any CPU hotplug opera
313 another way of preventing any callback
314 CPU, again preventing the rcuc/%u kthr
315 to do.
316
317 Name:
318 rcuop/%d and rcuos/%d
319
320 Purpose:
321 Offload RCU callbacks from the corresponding
322
323 To reduce its OS jitter, do at least one of th
324
325 1. Use affinity, cgroups, or other mechan
326 to execute on some other CPU.
327 2. Build with CONFIG_RCU_NOCB_CPU=n, whic
328 kthreads from being created in the fir
329 note that this will not eliminate OS j
330 shift it to RCU_SOFTIRQ.
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.