5 Copyright (C) 2004 BULL SA.
7 Written by Simon.Derr@bull.net
9 - Portions Copyright (c) 2004-2006 Silicon Graphics, Inc.
10 - Modified by Paul Jackson <pj@sgi.com>
11 - Modified by Christoph Lameter <cl@linux.com>
12 - Modified by Paul Menage <menage@google.com>
13 - Modified by Hidetoshi Seto <seto.hidetoshi@jp.fujitsu.com>
18 1.1 What are cpusets ?
19 1.2 Why are cpusets needed ?
20 1.3 How are cpusets implemented ?
21 1.4 What are exclusive cpusets ?
22 1.5 What is memory_pressure ?
23 1.6 What is memory spread ?
24 1.7 What is sched_load_balance ?
25 1.8 What is sched_relax_domain_level ?
26 1.9 How do I use cpusets ?
27 2. Usage Examples and Syntax
29 2.2 Adding/removing cpus
31 2.4 Attaching processes
38 1.1 What are cpusets ?
39 ----------------------
41 Cpusets provide a mechanism for assigning a set of CPUs and Memory
42 Nodes to a set of tasks. In this document "Memory Node" refers to
43 an on-line node that contains memory.
45 Cpusets constrain the CPU and Memory placement of tasks to only
46 the resources within a task's current cpuset. They form a nested
47 hierarchy visible in a virtual file system. These are the essential
48 hooks, beyond what is already present, required to manage dynamic
49 job placement on large systems.
51 Cpusets use the generic cgroup subsystem described in
52 Documentation/admin-guide/cgroup-v1/cgroups.rst.
54 Requests by a task, using the sched_setaffinity(2) system call to
55 include CPUs in its CPU affinity mask, and using the mbind(2) and
56 set_mempolicy(2) system calls to include Memory Nodes in its memory
57 policy, are both filtered through that task's cpuset, filtering out any
58 CPUs or Memory Nodes not in that cpuset. The scheduler will not
59 schedule a task on a CPU that is not allowed in its cpus_allowed
60 vector, and the kernel page allocator will not allocate a page on a
61 node that is not allowed in the requesting task's mems_allowed vector.
63 User level code may create and destroy cpusets by name in the cgroup
64 virtual file system, manage the attributes and permissions of these
65 cpusets and which CPUs and Memory Nodes are assigned to each cpuset,
66 specify and query to which cpuset a task is assigned, and list the
67 task pids assigned to a cpuset.
70 1.2 Why are cpusets needed ?
71 ----------------------------
73 The management of large computer systems, with many processors (CPUs),
74 complex memory cache hierarchies and multiple Memory Nodes having
75 non-uniform access times (NUMA) presents additional challenges for
76 the efficient scheduling and memory placement of processes.
78 Frequently more modest sized systems can be operated with adequate
79 efficiency just by letting the operating system automatically share
80 the available CPU and Memory resources amongst the requesting tasks.
82 But larger systems, which benefit more from careful processor and
83 memory placement to reduce memory access times and contention,
84 and which typically represent a larger investment for the customer,
85 can benefit from explicitly placing jobs on properly sized subsets of
88 This can be especially valuable on:
90 * Web Servers running multiple instances of the same web application,
91 * Servers running different applications (for instance, a web server
93 * NUMA systems running large HPC applications with demanding
94 performance characteristics.
96 These subsets, or "soft partitions" must be able to be dynamically
97 adjusted, as the job mix changes, without impacting other concurrently
98 executing jobs. The location of the running jobs pages may also be moved
99 when the memory locations are changed.
101 The kernel cpuset patch provides the minimum essential kernel
102 mechanisms required to efficiently implement such subsets. It
103 leverages existing CPU and Memory Placement facilities in the Linux
104 kernel to avoid any additional impact on the critical scheduler or
105 memory allocator code.
108 1.3 How are cpusets implemented ?
109 ---------------------------------
111 Cpusets provide a Linux kernel mechanism to constrain which CPUs and
112 Memory Nodes are used by a process or set of processes.
114 The Linux kernel already has a pair of mechanisms to specify on which
115 CPUs a task may be scheduled (sched_setaffinity) and on which Memory
116 Nodes it may obtain memory (mbind, set_mempolicy).
118 Cpusets extends these two mechanisms as follows:
120 - Cpusets are sets of allowed CPUs and Memory Nodes, known to the
122 - Each task in the system is attached to a cpuset, via a pointer
123 in the task structure to a reference counted cgroup structure.
124 - Calls to sched_setaffinity are filtered to just those CPUs
125 allowed in that task's cpuset.
126 - Calls to mbind and set_mempolicy are filtered to just
127 those Memory Nodes allowed in that task's cpuset.
128 - The root cpuset contains all the systems CPUs and Memory
130 - For any cpuset, one can define child cpusets containing a subset
131 of the parents CPU and Memory Node resources.
132 - The hierarchy of cpusets can be mounted at /dev/cpuset, for
133 browsing and manipulation from user space.
134 - A cpuset may be marked exclusive, which ensures that no other
135 cpuset (except direct ancestors and descendants) may contain
136 any overlapping CPUs or Memory Nodes.
137 - You can list all the tasks (by pid) attached to any cpuset.
139 The implementation of cpusets requires a few, simple hooks
140 into the rest of the kernel, none in performance critical paths:
142 - in init/main.c, to initialize the root cpuset at system boot.
143 - in fork and exit, to attach and detach a task from its cpuset.
144 - in sched_setaffinity, to mask the requested CPUs by what's
145 allowed in that task's cpuset.
146 - in sched.c migrate_live_tasks(), to keep migrating tasks within
147 the CPUs allowed by their cpuset, if possible.
148 - in the mbind and set_mempolicy system calls, to mask the requested
149 Memory Nodes by what's allowed in that task's cpuset.
150 - in page_alloc.c, to restrict memory to allowed nodes.
151 - in vmscan.c, to restrict page recovery to the current cpuset.
153 You should mount the "cgroup" filesystem type in order to enable
154 browsing and modifying the cpusets presently known to the kernel. No
155 new system calls are added for cpusets - all support for querying and
156 modifying cpusets is via this cpuset file system.
158 The /proc/<pid>/status file for each task has four added lines,
159 displaying the task's cpus_allowed (on which CPUs it may be scheduled)
160 and mems_allowed (on which Memory Nodes it may obtain memory),
161 in the two formats seen in the following example::
163 Cpus_allowed: ffffffff,ffffffff,ffffffff,ffffffff
164 Cpus_allowed_list: 0-127
165 Mems_allowed: ffffffff,ffffffff
166 Mems_allowed_list: 0-63
168 Each cpuset is represented by a directory in the cgroup file system
169 containing (on top of the standard cgroup files) the following
170 files describing that cpuset:
172 - cpuset.cpus: list of CPUs in that cpuset
173 - cpuset.mems: list of Memory Nodes in that cpuset
174 - cpuset.memory_migrate flag: if set, move pages to cpusets nodes
175 - cpuset.cpu_exclusive flag: is cpu placement exclusive?
176 - cpuset.mem_exclusive flag: is memory placement exclusive?
177 - cpuset.mem_hardwall flag: is memory allocation hardwalled
178 - cpuset.memory_pressure: measure of how much paging pressure in cpuset
179 - cpuset.memory_spread_page flag: if set, spread page cache evenly on allowed nodes
180 - cpuset.memory_spread_slab flag: if set, spread slab cache evenly on allowed nodes
181 - cpuset.sched_load_balance flag: if set, load balance within CPUs on that cpuset
182 - cpuset.sched_relax_domain_level: the searching range when migrating tasks
184 In addition, only the root cpuset has the following file:
186 - cpuset.memory_pressure_enabled flag: compute memory_pressure?
188 New cpusets are created using the mkdir system call or shell
189 command. The properties of a cpuset, such as its flags, allowed
190 CPUs and Memory Nodes, and attached tasks, are modified by writing
191 to the appropriate file in that cpusets directory, as listed above.
193 The named hierarchical structure of nested cpusets allows partitioning
194 a large system into nested, dynamically changeable, "soft-partitions".
196 The attachment of each task, automatically inherited at fork by any
197 children of that task, to a cpuset allows organizing the work load
198 on a system into related sets of tasks such that each set is constrained
199 to using the CPUs and Memory Nodes of a particular cpuset. A task
200 may be re-attached to any other cpuset, if allowed by the permissions
201 on the necessary cpuset file system directories.
203 Such management of a system "in the large" integrates smoothly with
204 the detailed placement done on individual tasks and memory regions
205 using the sched_setaffinity, mbind and set_mempolicy system calls.
207 The following rules apply to each cpuset:
209 - Its CPUs and Memory Nodes must be a subset of its parents.
210 - It can't be marked exclusive unless its parent is.
211 - If its cpu or memory is exclusive, they may not overlap any sibling.
213 These rules, and the natural hierarchy of cpusets, enable efficient
214 enforcement of the exclusive guarantee, without having to scan all
215 cpusets every time any of them change to ensure nothing overlaps a
216 exclusive cpuset. Also, the use of a Linux virtual file system (vfs)
217 to represent the cpuset hierarchy provides for a familiar permission
218 and name space for cpusets, with a minimum of additional kernel code.
220 The cpus and mems files in the root (top_cpuset) cpuset are
221 read-only. The cpus file automatically tracks the value of
222 cpu_online_mask using a CPU hotplug notifier, and the mems file
223 automatically tracks the value of node_states[N_MEMORY]--i.e.,
224 nodes with memory--using the cpuset_track_online_nodes() hook.
227 1.4 What are exclusive cpusets ?
228 --------------------------------
230 If a cpuset is cpu or mem exclusive, no other cpuset, other than
231 a direct ancestor or descendant, may share any of the same CPUs or
234 A cpuset that is cpuset.mem_exclusive *or* cpuset.mem_hardwall is "hardwalled",
235 i.e. it restricts kernel allocations for page, buffer and other data
236 commonly shared by the kernel across multiple users. All cpusets,
237 whether hardwalled or not, restrict allocations of memory for user
238 space. This enables configuring a system so that several independent
239 jobs can share common kernel data, such as file system pages, while
240 isolating each job's user allocation in its own cpuset. To do this,
241 construct a large mem_exclusive cpuset to hold all the jobs, and
242 construct child, non-mem_exclusive cpusets for each individual job.
243 Only a small amount of typical kernel memory, such as requests from
244 interrupt handlers, is allowed to be taken outside even a
245 mem_exclusive cpuset.
248 1.5 What is memory_pressure ?
249 -----------------------------
250 The memory_pressure of a cpuset provides a simple per-cpuset metric
251 of the rate that the tasks in a cpuset are attempting to free up in
252 use memory on the nodes of the cpuset to satisfy additional memory
255 This enables batch managers monitoring jobs running in dedicated
256 cpusets to efficiently detect what level of memory pressure that job
259 This is useful both on tightly managed systems running a wide mix of
260 submitted jobs, which may choose to terminate or re-prioritize jobs that
261 are trying to use more memory than allowed on the nodes assigned to them,
262 and with tightly coupled, long running, massively parallel scientific
263 computing jobs that will dramatically fail to meet required performance
264 goals if they start to use more memory than allowed to them.
266 This mechanism provides a very economical way for the batch manager
267 to monitor a cpuset for signs of memory pressure. It's up to the
268 batch manager or other user code to decide what to do about it and
272 Unless this feature is enabled by writing "1" to the special file
273 /dev/cpuset/memory_pressure_enabled, the hook in the rebalance
274 code of __alloc_pages() for this metric reduces to simply noticing
275 that the cpuset_memory_pressure_enabled flag is zero. So only
276 systems that enable this feature will compute the metric.
278 Why a per-cpuset, running average:
280 Because this meter is per-cpuset, rather than per-task or mm,
281 the system load imposed by a batch scheduler monitoring this
282 metric is sharply reduced on large systems, because a scan of
283 the tasklist can be avoided on each set of queries.
285 Because this meter is a running average, instead of an accumulating
286 counter, a batch scheduler can detect memory pressure with a
287 single read, instead of having to read and accumulate results
288 for a period of time.
290 Because this meter is per-cpuset rather than per-task or mm,
291 the batch scheduler can obtain the key information, memory
292 pressure in a cpuset, with a single read, rather than having to
293 query and accumulate results over all the (dynamically changing)
294 set of tasks in the cpuset.
296 A per-cpuset simple digital filter (requires a spinlock and 3 words
297 of data per-cpuset) is kept, and updated by any task attached to that
298 cpuset, if it enters the synchronous (direct) page reclaim code.
300 A per-cpuset file provides an integer number representing the recent
301 (half-life of 10 seconds) rate of direct page reclaims caused by
302 the tasks in the cpuset, in units of reclaims attempted per second,
306 1.6 What is memory spread ?
307 ---------------------------
308 There are two boolean flag files per cpuset that control where the
309 kernel allocates pages for the file system buffers and related in
310 kernel data structures. They are called 'cpuset.memory_spread_page' and
311 'cpuset.memory_spread_slab'.
313 If the per-cpuset boolean flag file 'cpuset.memory_spread_page' is set, then
314 the kernel will spread the file system buffers (page cache) evenly
315 over all the nodes that the faulting task is allowed to use, instead
316 of preferring to put those pages on the node where the task is running.
318 If the per-cpuset boolean flag file 'cpuset.memory_spread_slab' is set,
319 then the kernel will spread some file system related slab caches,
320 such as for inodes and dentries evenly over all the nodes that the
321 faulting task is allowed to use, instead of preferring to put those
322 pages on the node where the task is running.
324 The setting of these flags does not affect anonymous data segment or
325 stack segment pages of a task.
327 By default, both kinds of memory spreading are off, and memory
328 pages are allocated on the node local to where the task is running,
329 except perhaps as modified by the task's NUMA mempolicy or cpuset
330 configuration, so long as sufficient free memory pages are available.
332 When new cpusets are created, they inherit the memory spread settings
335 Setting memory spreading causes allocations for the affected page
336 or slab caches to ignore the task's NUMA mempolicy and be spread
337 instead. Tasks using mbind() or set_mempolicy() calls to set NUMA
338 mempolicies will not notice any change in these calls as a result of
339 their containing task's memory spread settings. If memory spreading
340 is turned off, then the currently specified NUMA mempolicy once again
341 applies to memory page allocations.
343 Both 'cpuset.memory_spread_page' and 'cpuset.memory_spread_slab' are boolean flag
344 files. By default they contain "0", meaning that the feature is off
345 for that cpuset. If a "1" is written to that file, then that turns
346 the named feature on.
348 The implementation is simple.
350 Setting the flag 'cpuset.memory_spread_page' turns on a per-process flag
351 PFA_SPREAD_PAGE for each task that is in that cpuset or subsequently
352 joins that cpuset. The page allocation calls for the page cache
353 is modified to perform an inline check for this PFA_SPREAD_PAGE task
354 flag, and if set, a call to a new routine cpuset_mem_spread_node()
355 returns the node to prefer for the allocation.
357 Similarly, setting 'cpuset.memory_spread_slab' turns on the flag
358 PFA_SPREAD_SLAB, and appropriately marked slab caches will allocate
359 pages from the node returned by cpuset_mem_spread_node().
361 The cpuset_mem_spread_node() routine is also simple. It uses the
362 value of a per-task rotor cpuset_mem_spread_rotor to select the next
363 node in the current task's mems_allowed to prefer for the allocation.
365 This memory placement policy is also known (in other contexts) as
366 round-robin or interleave.
368 This policy can provide substantial improvements for jobs that need
369 to place thread local data on the corresponding node, but that need
370 to access large file system data sets that need to be spread across
371 the several nodes in the jobs cpuset in order to fit. Without this
372 policy, especially for jobs that might have one thread reading in the
373 data set, the memory allocation across the nodes in the jobs cpuset
374 can become very uneven.
376 1.7 What is sched_load_balance ?
377 --------------------------------
379 The kernel scheduler (kernel/sched/core.c) automatically load balances
380 tasks. If one CPU is underutilized, kernel code running on that
381 CPU will look for tasks on other more overloaded CPUs and move those
382 tasks to itself, within the constraints of such placement mechanisms
383 as cpusets and sched_setaffinity.
385 The algorithmic cost of load balancing and its impact on key shared
386 kernel data structures such as the task list increases more than
387 linearly with the number of CPUs being balanced. So the scheduler
388 has support to partition the systems CPUs into a number of sched
389 domains such that it only load balances within each sched domain.
390 Each sched domain covers some subset of the CPUs in the system;
391 no two sched domains overlap; some CPUs might not be in any sched
392 domain and hence won't be load balanced.
394 Put simply, it costs less to balance between two smaller sched domains
395 than one big one, but doing so means that overloads in one of the
396 two domains won't be load balanced to the other one.
398 By default, there is one sched domain covering all CPUs, including those
399 marked isolated using the kernel boot time "isolcpus=" argument. However,
400 the isolated CPUs will not participate in load balancing, and will not
401 have tasks running on them unless explicitly assigned.
403 This default load balancing across all CPUs is not well suited for
404 the following two situations:
406 1) On large systems, load balancing across many CPUs is expensive.
407 If the system is managed using cpusets to place independent jobs
408 on separate sets of CPUs, full load balancing is unnecessary.
409 2) Systems supporting realtime on some CPUs need to minimize
410 system overhead on those CPUs, including avoiding task load
411 balancing if that is not needed.
413 When the per-cpuset flag "cpuset.sched_load_balance" is enabled (the default
414 setting), it requests that all the CPUs in that cpusets allowed 'cpuset.cpus'
415 be contained in a single sched domain, ensuring that load balancing
416 can move a task (not otherwised pinned, as by sched_setaffinity)
417 from any CPU in that cpuset to any other.
419 When the per-cpuset flag "cpuset.sched_load_balance" is disabled, then the
420 scheduler will avoid load balancing across the CPUs in that cpuset,
421 --except-- in so far as is necessary because some overlapping cpuset
422 has "sched_load_balance" enabled.
424 So, for example, if the top cpuset has the flag "cpuset.sched_load_balance"
425 enabled, then the scheduler will have one sched domain covering all
426 CPUs, and the setting of the "cpuset.sched_load_balance" flag in any other
427 cpusets won't matter, as we're already fully load balancing.
429 Therefore in the above two situations, the top cpuset flag
430 "cpuset.sched_load_balance" should be disabled, and only some of the smaller,
431 child cpusets have this flag enabled.
433 When doing this, you don't usually want to leave any unpinned tasks in
434 the top cpuset that might use non-trivial amounts of CPU, as such tasks
435 may be artificially constrained to some subset of CPUs, depending on
436 the particulars of this flag setting in descendant cpusets. Even if
437 such a task could use spare CPU cycles in some other CPUs, the kernel
438 scheduler might not consider the possibility of load balancing that
439 task to that underused CPU.
441 Of course, tasks pinned to a particular CPU can be left in a cpuset
442 that disables "cpuset.sched_load_balance" as those tasks aren't going anywhere
445 There is an impedance mismatch here, between cpusets and sched domains.
446 Cpusets are hierarchical and nest. Sched domains are flat; they don't
447 overlap and each CPU is in at most one sched domain.
449 It is necessary for sched domains to be flat because load balancing
450 across partially overlapping sets of CPUs would risk unstable dynamics
451 that would be beyond our understanding. So if each of two partially
452 overlapping cpusets enables the flag 'cpuset.sched_load_balance', then we
453 form a single sched domain that is a superset of both. We won't move
454 a task to a CPU outside its cpuset, but the scheduler load balancing
455 code might waste some compute cycles considering that possibility.
457 This mismatch is why there is not a simple one-to-one relation
458 between which cpusets have the flag "cpuset.sched_load_balance" enabled,
459 and the sched domain configuration. If a cpuset enables the flag, it
460 will get balancing across all its CPUs, but if it disables the flag,
461 it will only be assured of no load balancing if no other overlapping
462 cpuset enables the flag.
464 If two cpusets have partially overlapping 'cpuset.cpus' allowed, and only
465 one of them has this flag enabled, then the other may find its
466 tasks only partially load balanced, just on the overlapping CPUs.
467 This is just the general case of the top_cpuset example given a few
468 paragraphs above. In the general case, as in the top cpuset case,
469 don't leave tasks that might use non-trivial amounts of CPU in
470 such partially load balanced cpusets, as they may be artificially
471 constrained to some subset of the CPUs allowed to them, for lack of
472 load balancing to the other CPUs.
474 CPUs in "cpuset.isolcpus" were excluded from load balancing by the
475 isolcpus= kernel boot option, and will never be load balanced regardless
476 of the value of "cpuset.sched_load_balance" in any cpuset.
478 1.7.1 sched_load_balance implementation details.
479 ------------------------------------------------
481 The per-cpuset flag 'cpuset.sched_load_balance' defaults to enabled (contrary
482 to most cpuset flags.) When enabled for a cpuset, the kernel will
483 ensure that it can load balance across all the CPUs in that cpuset
484 (makes sure that all the CPUs in the cpus_allowed of that cpuset are
485 in the same sched domain.)
487 If two overlapping cpusets both have 'cpuset.sched_load_balance' enabled,
488 then they will be (must be) both in the same sched domain.
490 If, as is the default, the top cpuset has 'cpuset.sched_load_balance' enabled,
491 then by the above that means there is a single sched domain covering
492 the whole system, regardless of any other cpuset settings.
494 The kernel commits to user space that it will avoid load balancing
495 where it can. It will pick as fine a granularity partition of sched
496 domains as it can while still providing load balancing for any set
497 of CPUs allowed to a cpuset having 'cpuset.sched_load_balance' enabled.
499 The internal kernel cpuset to scheduler interface passes from the
500 cpuset code to the scheduler code a partition of the load balanced
501 CPUs in the system. This partition is a set of subsets (represented
502 as an array of struct cpumask) of CPUs, pairwise disjoint, that cover
503 all the CPUs that must be load balanced.
505 The cpuset code builds a new such partition and passes it to the
506 scheduler sched domain setup code, to have the sched domains rebuilt
507 as necessary, whenever:
509 - the 'cpuset.sched_load_balance' flag of a cpuset with non-empty CPUs changes,
510 - or CPUs come or go from a cpuset with this flag enabled,
511 - or 'cpuset.sched_relax_domain_level' value of a cpuset with non-empty CPUs
512 and with this flag enabled changes,
513 - or a cpuset with non-empty CPUs and with this flag enabled is removed,
514 - or a cpu is offlined/onlined.
516 This partition exactly defines what sched domains the scheduler should
517 setup - one sched domain for each element (struct cpumask) in the
520 The scheduler remembers the currently active sched domain partitions.
521 When the scheduler routine partition_sched_domains() is invoked from
522 the cpuset code to update these sched domains, it compares the new
523 partition requested with the current, and updates its sched domains,
524 removing the old and adding the new, for each change.
527 1.8 What is sched_relax_domain_level ?
528 --------------------------------------
530 In sched domain, the scheduler migrates tasks in 2 ways; periodic load
531 balance on tick, and at time of some schedule events.
533 When a task is woken up, scheduler try to move the task on idle CPU.
534 For example, if a task A running on CPU X activates another task B
535 on the same CPU X, and if CPU Y is X's sibling and performing idle,
536 then scheduler migrate task B to CPU Y so that task B can start on
537 CPU Y without waiting task A on CPU X.
539 And if a CPU run out of tasks in its runqueue, the CPU try to pull
540 extra tasks from other busy CPUs to help them before it is going to
543 Of course it takes some searching cost to find movable tasks and/or
544 idle CPUs, the scheduler might not search all CPUs in the domain
545 every time. In fact, in some architectures, the searching ranges on
546 events are limited in the same socket or node where the CPU locates,
547 while the load balance on tick searches all.
549 For example, assume CPU Z is relatively far from CPU X. Even if CPU Z
550 is idle while CPU X and the siblings are busy, scheduler can't migrate
551 woken task B from X to Z since it is out of its searching range.
552 As the result, task B on CPU X need to wait task A or wait load balance
553 on the next tick. For some applications in special situation, waiting
554 1 tick may be too long.
556 The 'cpuset.sched_relax_domain_level' file allows you to request changing
557 this searching range as you like. This file takes int value which
558 indicates size of searching range in levels ideally as follows,
559 otherwise initial value -1 that indicates the cpuset has no request.
561 ====== ===========================================================
562 -1 no request. use system default or follow request of others.
564 1 search siblings (hyperthreads in a core).
565 2 search cores in a package.
566 3 search cpus in a node [= system wide on non-NUMA system]
567 4 search nodes in a chunk of node [on NUMA system]
568 5 search system wide [on NUMA system]
569 ====== ===========================================================
571 The system default is architecture dependent. The system default
572 can be changed using the relax_domain_level= boot parameter.
574 This file is per-cpuset and affect the sched domain where the cpuset
575 belongs to. Therefore if the flag 'cpuset.sched_load_balance' of a cpuset
576 is disabled, then 'cpuset.sched_relax_domain_level' have no effect since
577 there is no sched domain belonging the cpuset.
579 If multiple cpusets are overlapping and hence they form a single sched
580 domain, the largest value among those is used. Be careful, if one
581 requests 0 and others are -1 then 0 is used.
583 Note that modifying this file will have both good and bad effects,
584 and whether it is acceptable or not depends on your situation.
585 Don't modify this file if you are not sure.
587 If your situation is:
589 - The migration costs between each cpu can be assumed considerably
590 small(for you) due to your special application's behavior or
591 special hardware support for CPU cache etc.
592 - The searching cost doesn't have impact(for you) or you can make
593 the searching cost enough small by managing cpuset to compact etc.
594 - The latency is required even it sacrifices cache hit rate etc.
595 then increasing 'sched_relax_domain_level' would benefit you.
598 1.9 How do I use cpusets ?
599 --------------------------
601 In order to minimize the impact of cpusets on critical kernel
602 code, such as the scheduler, and due to the fact that the kernel
603 does not support one task updating the memory placement of another
604 task directly, the impact on a task of changing its cpuset CPU
605 or Memory Node placement, or of changing to which cpuset a task
606 is attached, is subtle.
608 If a cpuset has its Memory Nodes modified, then for each task attached
609 to that cpuset, the next time that the kernel attempts to allocate
610 a page of memory for that task, the kernel will notice the change
611 in the task's cpuset, and update its per-task memory placement to
612 remain within the new cpusets memory placement. If the task was using
613 mempolicy MPOL_BIND, and the nodes to which it was bound overlap with
614 its new cpuset, then the task will continue to use whatever subset
615 of MPOL_BIND nodes are still allowed in the new cpuset. If the task
616 was using MPOL_BIND and now none of its MPOL_BIND nodes are allowed
617 in the new cpuset, then the task will be essentially treated as if it
618 was MPOL_BIND bound to the new cpuset (even though its NUMA placement,
619 as queried by get_mempolicy(), doesn't change). If a task is moved
620 from one cpuset to another, then the kernel will adjust the task's
621 memory placement, as above, the next time that the kernel attempts
622 to allocate a page of memory for that task.
624 If a cpuset has its 'cpuset.cpus' modified, then each task in that cpuset
625 will have its allowed CPU placement changed immediately. Similarly,
626 if a task's pid is written to another cpuset's 'tasks' file, then its
627 allowed CPU placement is changed immediately. If such a task had been
628 bound to some subset of its cpuset using the sched_setaffinity() call,
629 the task will be allowed to run on any CPU allowed in its new cpuset,
630 negating the effect of the prior sched_setaffinity() call.
632 In summary, the memory placement of a task whose cpuset is changed is
633 updated by the kernel, on the next allocation of a page for that task,
634 and the processor placement is updated immediately.
636 Normally, once a page is allocated (given a physical page
637 of main memory) then that page stays on whatever node it
638 was allocated, so long as it remains allocated, even if the
639 cpusets memory placement policy 'cpuset.mems' subsequently changes.
640 If the cpuset flag file 'cpuset.memory_migrate' is set true, then when
641 tasks are attached to that cpuset, any pages that task had
642 allocated to it on nodes in its previous cpuset are migrated
643 to the task's new cpuset. The relative placement of the page within
644 the cpuset is preserved during these migration operations if possible.
645 For example if the page was on the second valid node of the prior cpuset
646 then the page will be placed on the second valid node of the new cpuset.
648 Also if 'cpuset.memory_migrate' is set true, then if that cpuset's
649 'cpuset.mems' file is modified, pages allocated to tasks in that
650 cpuset, that were on nodes in the previous setting of 'cpuset.mems',
651 will be moved to nodes in the new setting of 'mems.'
652 Pages that were not in the task's prior cpuset, or in the cpuset's
653 prior 'cpuset.mems' setting, will not be moved.
655 There is an exception to the above. If hotplug functionality is used
656 to remove all the CPUs that are currently assigned to a cpuset,
657 then all the tasks in that cpuset will be moved to the nearest ancestor
658 with non-empty cpus. But the moving of some (or all) tasks might fail if
659 cpuset is bound with another cgroup subsystem which has some restrictions
660 on task attaching. In this failing case, those tasks will stay
661 in the original cpuset, and the kernel will automatically update
662 their cpus_allowed to allow all online CPUs. When memory hotplug
663 functionality for removing Memory Nodes is available, a similar exception
664 is expected to apply there as well. In general, the kernel prefers to
665 violate cpuset placement, over starving a task that has had all
666 its allowed CPUs or Memory Nodes taken offline.
668 There is a second exception to the above. GFP_ATOMIC requests are
669 kernel internal allocations that must be satisfied, immediately.
670 The kernel may drop some request, in rare cases even panic, if a
671 GFP_ATOMIC alloc fails. If the request cannot be satisfied within
672 the current task's cpuset, then we relax the cpuset, and look for
673 memory anywhere we can find it. It's better to violate the cpuset
674 than stress the kernel.
676 To start a new job that is to be contained within a cpuset, the steps are:
678 1) mkdir /sys/fs/cgroup/cpuset
679 2) mount -t cgroup -ocpuset cpuset /sys/fs/cgroup/cpuset
680 3) Create the new cpuset by doing mkdir's and write's (or echo's) in
681 the /sys/fs/cgroup/cpuset virtual file system.
682 4) Start a task that will be the "founding father" of the new job.
683 5) Attach that task to the new cpuset by writing its pid to the
684 /sys/fs/cgroup/cpuset tasks file for that cpuset.
685 6) fork, exec or clone the job tasks from this founding father task.
687 For example, the following sequence of commands will setup a cpuset
688 named "Charlie", containing just CPUs 2 and 3, and Memory Node 1,
689 and then start a subshell 'sh' in that cpuset::
691 mount -t cgroup -ocpuset cpuset /sys/fs/cgroup/cpuset
692 cd /sys/fs/cgroup/cpuset
695 /bin/echo 2-3 > cpuset.cpus
696 /bin/echo 1 > cpuset.mems
699 # The subshell 'sh' is now running in cpuset Charlie
700 # The next line should display '/Charlie'
701 cat /proc/self/cpuset
703 There are ways to query or modify cpusets:
705 - via the cpuset file system directly, using the various cd, mkdir, echo,
706 cat, rmdir commands from the shell, or their equivalent from C.
707 - via the C library libcpuset.
708 - via the C library libcgroup.
709 (http://sourceforge.net/projects/libcg/)
710 - via the python application cset.
711 (http://code.google.com/p/cpuset/)
713 The sched_setaffinity calls can also be done at the shell prompt using
714 SGI's runon or Robert Love's taskset. The mbind and set_mempolicy
715 calls can be done at the shell prompt using the numactl command
716 (part of Andi Kleen's numa package).
718 2. Usage Examples and Syntax
719 ============================
724 Creating, modifying, using the cpusets can be done through the cpuset
728 # mount -t cgroup -o cpuset cpuset /sys/fs/cgroup/cpuset
730 Then under /sys/fs/cgroup/cpuset you can find a tree that corresponds to the
731 tree of the cpusets in the system. For instance, /sys/fs/cgroup/cpuset
732 is the cpuset that holds the whole system.
734 If you want to create a new cpuset under /sys/fs/cgroup/cpuset::
736 # cd /sys/fs/cgroup/cpuset
739 Now you want to do something with this cpuset::
743 In this directory you can find several files::
746 cgroup.clone_children cpuset.memory_pressure
747 cgroup.event_control cpuset.memory_spread_page
748 cgroup.procs cpuset.memory_spread_slab
749 cpuset.cpu_exclusive cpuset.mems
750 cpuset.cpus cpuset.sched_load_balance
751 cpuset.mem_exclusive cpuset.sched_relax_domain_level
752 cpuset.mem_hardwall notify_on_release
753 cpuset.memory_migrate tasks
755 Reading them will give you information about the state of this cpuset:
756 the CPUs and Memory Nodes it can use, the processes that are using
757 it, its properties. By writing to these files you can manipulate
762 # /bin/echo 1 > cpuset.cpu_exclusive
766 # /bin/echo 0-7 > cpuset.cpus
770 # /bin/echo 0-7 > cpuset.mems
772 Now attach your shell to this cpuset::
774 # /bin/echo $$ > tasks
776 You can also create cpusets inside your cpuset by using mkdir in this
781 To remove a cpuset, just use rmdir::
785 This will fail if the cpuset is in use (has cpusets inside, or has
788 Note that for legacy reasons, the "cpuset" filesystem exists as a
789 wrapper around the cgroup filesystem.
793 mount -t cpuset X /sys/fs/cgroup/cpuset
797 mount -t cgroup -ocpuset,noprefix X /sys/fs/cgroup/cpuset
798 echo "/sbin/cpuset_release_agent" > /sys/fs/cgroup/cpuset/release_agent
800 2.2 Adding/removing cpus
801 ------------------------
803 This is the syntax to use when writing in the cpus or mems files
804 in cpuset directories::
806 # /bin/echo 1-4 > cpuset.cpus -> set cpus list to cpus 1,2,3,4
807 # /bin/echo 1,2,3,4 > cpuset.cpus -> set cpus list to cpus 1,2,3,4
809 To add a CPU to a cpuset, write the new list of CPUs including the
810 CPU to be added. To add 6 to the above cpuset::
812 # /bin/echo 1-4,6 > cpuset.cpus -> set cpus list to cpus 1,2,3,4,6
814 Similarly to remove a CPU from a cpuset, write the new list of CPUs
815 without the CPU to be removed.
817 To remove all the CPUs::
819 # /bin/echo "" > cpuset.cpus -> clear cpus list
824 The syntax is very simple::
826 # /bin/echo 1 > cpuset.cpu_exclusive -> set flag 'cpuset.cpu_exclusive'
827 # /bin/echo 0 > cpuset.cpu_exclusive -> unset flag 'cpuset.cpu_exclusive'
829 2.4 Attaching processes
830 -----------------------
834 # /bin/echo PID > tasks
836 Note that it is PID, not PIDs. You can only attach ONE task at a time.
837 If you have several tasks to attach, you have to do it one after another::
839 # /bin/echo PID1 > tasks
840 # /bin/echo PID2 > tasks
842 # /bin/echo PIDn > tasks
849 what's up with this '/bin/echo' ?
852 bash's builtin 'echo' command does not check calls to write() against
853 errors. If you use it in the cpuset file system, you won't be
854 able to tell whether a command succeeded or failed.
857 When I attach processes, only the first of the line gets really attached !
860 We can only return one error code per call to write(). So you should also
866 Web: http://www.bullopensource.org/cpuset