Root/Documentation/kernel-per-CPU-kthreads.txt

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. Do any of the following needed to avoid jitter that your
166    application cannot tolerate:
167    a. Build your kernel with CONFIG_SLUB=y rather than
168        CONFIG_SLAB=y, thus avoiding the slab allocator's periodic
169        use of each CPU's workqueues to run its cache_reap()
170        function.
171    b. Avoid using oprofile, thus avoiding OS jitter from
172        wq_sync_buffer().
173    c. Limit your CPU frequency so that a CPU-frequency
174        governor is not required, possibly enlisting the aid of
175        special heatsinks or other cooling technologies. If done
176        correctly, and if you CPU architecture permits, you should
177        be able to build your kernel with CONFIG_CPU_FREQ=n to
178        avoid the CPU-frequency governor periodically running
179        on each CPU, including cs_dbs_timer() and od_dbs_timer().
180        WARNING: Please check your CPU specifications to
181        make sure that this is safe on your particular system.
182    d. It is not possible to entirely get rid of OS jitter
183        from vmstat_update() on CONFIG_SMP=y systems, but you
184        can decrease its frequency by writing a large value to
185        /proc/sys/vm/stat_interval. The default value is HZ,
186        for an interval of one second. Of course, larger values
187        will make your virtual-memory statistics update more
188        slowly. Of course, you can also run your workload at
189        a real-time priority, thus preempting vmstat_update().
190    e. If running on high-end powerpc servers, build with
191        CONFIG_PPC_RTAS_DAEMON=n. This prevents the RTAS
192        daemon from running on each CPU every second or so.
193        (This will require editing Kconfig files and will defeat
194        this platform's RAS functionality.) This avoids jitter
195        due to the rtas_event_scan() function.
196        WARNING: Please check your CPU specifications to
197        make sure that this is safe on your particular system.
198    f. If running on Cell Processor, build your kernel with
199        CBE_CPUFREQ_SPU_GOVERNOR=n to avoid OS jitter from
200        spu_gov_work().
201        WARNING: Please check your CPU specifications to
202        make sure that this is safe on your particular system.
203    g. If running on PowerMAC, build your kernel with
204        CONFIG_PMAC_RACKMETER=n to disable the CPU-meter,
205        avoiding OS jitter from rackmeter_do_timer().
206
207Name: rcuc/%u
208Purpose: Execute RCU callbacks in CONFIG_RCU_BOOST=y kernels.
209To reduce its OS jitter, do at least one of the following:
2101. Build the kernel with CONFIG_PREEMPT=n. This prevents these
211    kthreads from being created in the first place, and also obviates
212    the need for RCU priority boosting. This approach is feasible
213    for workloads that do not require high degrees of responsiveness.
2142. Build the kernel with CONFIG_RCU_BOOST=n. This prevents these
215    kthreads from being created in the first place. This approach
216    is feasible only if your workload never requires RCU priority
217    boosting, for example, if you ensure frequent idle time on all
218    CPUs that might execute within the kernel.
2193. Build with CONFIG_RCU_NOCB_CPU=y and CONFIG_RCU_NOCB_CPU_ALL=y,
220    which offloads all RCU callbacks to kthreads that can be moved
221    off of CPUs susceptible to OS jitter. This approach prevents the
222    rcuc/%u kthreads from having any work to do, so that they are
223    never awakened.
2244. Ensure that the CPU never enters the kernel, and, in particular,
225    avoid initiating any CPU hotplug operations on this CPU. This is
226    another way of preventing any callbacks from being queued on the
227    CPU, again preventing the rcuc/%u kthreads from having any work
228    to do.
229
230Name: rcuob/%d, rcuop/%d, and rcuos/%d
231Purpose: Offload RCU callbacks from the corresponding CPU.
232To reduce its OS jitter, do at least one of the following:
2331. Use affinity, cgroups, or other mechanism to force these kthreads
234    to execute on some other CPU.
2352. Build with CONFIG_RCU_NOCB_CPU=n, which will prevent these
236    kthreads from being created in the first place. However, please
237    note that this will not eliminate OS jitter, but will instead
238    shift it to RCU_SOFTIRQ.
239
240Name: watchdog/%u
241Purpose: Detect software lockups on each CPU.
242To reduce its OS jitter, do at least one of the following:
2431. Build with CONFIG_LOCKUP_DETECTOR=n, which will prevent these
244    kthreads from being created in the first place.
2452. Echo a zero to /proc/sys/kernel/watchdog to disable the
246    watchdog timer.
2473. Echo a large number of /proc/sys/kernel/watchdog_thresh in
248    order to reduce the frequency of OS jitter due to the watchdog
249    timer down to a level that is acceptable for your workload.
250

Archive Download this file



interactive