1.. _cpusets:
2
3=======
4CPUSETS
5=======
6
7Copyright (C) 2004 BULL SA.
8
9Written by Simon.Derr@bull.net
10
11- Portions Copyright (c) 2004-2006 Silicon Graphics, Inc.
12- Modified by Paul Jackson <pj@sgi.com>
13- Modified by Christoph Lameter <cl@linux.com>
14- Modified by Paul Menage <menage@google.com>
15- Modified by Hidetoshi Seto <seto.hidetoshi@jp.fujitsu.com>
16
17.. CONTENTS:
18
19   1. Cpusets
20     1.1 What are cpusets ?
21     1.2 Why are cpusets needed ?
22     1.3 How are cpusets implemented ?
23     1.4 What are exclusive cpusets ?
24     1.5 What is memory_pressure ?
25     1.6 What is memory spread ?
26     1.7 What is sched_load_balance ?
27     1.8 What is sched_relax_domain_level ?
28     1.9 How do I use cpusets ?
29   2. Usage Examples and Syntax
30     2.1 Basic Usage
31     2.2 Adding/removing cpus
32     2.3 Setting flags
33     2.4 Attaching processes
34   3. Questions
35   4. Contact
36
371. Cpusets
38==========
39
401.1 What are cpusets ?
41----------------------
42
43Cpusets provide a mechanism for assigning a set of CPUs and Memory
44Nodes to a set of tasks.   In this document "Memory Node" refers to
45an on-line node that contains memory.
46
47Cpusets constrain the CPU and Memory placement of tasks to only
48the resources within a task's current cpuset.  They form a nested
49hierarchy visible in a virtual file system.  These are the essential
50hooks, beyond what is already present, required to manage dynamic
51job placement on large systems.
52
53Cpusets use the generic cgroup subsystem described in
54Documentation/admin-guide/cgroup-v1/cgroups.rst.
55
56Requests by a task, using the sched_setaffinity(2) system call to
57include CPUs in its CPU affinity mask, and using the mbind(2) and
58set_mempolicy(2) system calls to include Memory Nodes in its memory
59policy, are both filtered through that task's cpuset, filtering out any
60CPUs or Memory Nodes not in that cpuset.  The scheduler will not
61schedule a task on a CPU that is not allowed in its cpus_allowed
62vector, and the kernel page allocator will not allocate a page on a
63node that is not allowed in the requesting task's mems_allowed vector.
64
65User level code may create and destroy cpusets by name in the cgroup
66virtual file system, manage the attributes and permissions of these
67cpusets and which CPUs and Memory Nodes are assigned to each cpuset,
68specify and query to which cpuset a task is assigned, and list the
69task pids assigned to a cpuset.
70
71
721.2 Why are cpusets needed ?
73----------------------------
74
75The management of large computer systems, with many processors (CPUs),
76complex memory cache hierarchies and multiple Memory Nodes having
77non-uniform access times (NUMA) presents additional challenges for
78the efficient scheduling and memory placement of processes.
79
80Frequently more modest sized systems can be operated with adequate
81efficiency just by letting the operating system automatically share
82the available CPU and Memory resources amongst the requesting tasks.
83
84But larger systems, which benefit more from careful processor and
85memory placement to reduce memory access times and contention,
86and which typically represent a larger investment for the customer,
87can benefit from explicitly placing jobs on properly sized subsets of
88the system.
89
90This can be especially valuable on:
91
92    * Web Servers running multiple instances of the same web application,
93    * Servers running different applications (for instance, a web server
94      and a database), or
95    * NUMA systems running large HPC applications with demanding
96      performance characteristics.
97
98These subsets, or "soft partitions" must be able to be dynamically
99adjusted, as the job mix changes, without impacting other concurrently
100executing jobs. The location of the running jobs pages may also be moved
101when the memory locations are changed.
102
103The kernel cpuset patch provides the minimum essential kernel
104mechanisms required to efficiently implement such subsets.  It
105leverages existing CPU and Memory Placement facilities in the Linux
106kernel to avoid any additional impact on the critical scheduler or
107memory allocator code.
108
109
1101.3 How are cpusets implemented ?
111---------------------------------
112
113Cpusets provide a Linux kernel mechanism to constrain which CPUs and
114Memory Nodes are used by a process or set of processes.
115
116The Linux kernel already has a pair of mechanisms to specify on which
117CPUs a task may be scheduled (sched_setaffinity) and on which Memory
118Nodes it may obtain memory (mbind, set_mempolicy).
119
120Cpusets extends these two mechanisms as follows:
121
122 - Cpusets are sets of allowed CPUs and Memory Nodes, known to the
123   kernel.
124 - Each task in the system is attached to a cpuset, via a pointer
125   in the task structure to a reference counted cgroup structure.
126 - Calls to sched_setaffinity are filtered to just those CPUs
127   allowed in that task's cpuset.
128 - Calls to mbind and set_mempolicy are filtered to just
129   those Memory Nodes allowed in that task's cpuset.
130 - The root cpuset contains all the systems CPUs and Memory
131   Nodes.
132 - For any cpuset, one can define child cpusets containing a subset
133   of the parents CPU and Memory Node resources.
134 - The hierarchy of cpusets can be mounted at /dev/cpuset, for
135   browsing and manipulation from user space.
136 - A cpuset may be marked exclusive, which ensures that no other
137   cpuset (except direct ancestors and descendants) may contain
138   any overlapping CPUs or Memory Nodes.
139 - You can list all the tasks (by pid) attached to any cpuset.
140
141The implementation of cpusets requires a few, simple hooks
142into the rest of the kernel, none in performance critical paths:
143
144 - in init/main.c, to initialize the root cpuset at system boot.
145 - in fork and exit, to attach and detach a task from its cpuset.
146 - in sched_setaffinity, to mask the requested CPUs by what's
147   allowed in that task's cpuset.
148 - in sched.c migrate_live_tasks(), to keep migrating tasks within
149   the CPUs allowed by their cpuset, if possible.
150 - in the mbind and set_mempolicy system calls, to mask the requested
151   Memory Nodes by what's allowed in that task's cpuset.
152 - in page_alloc.c, to restrict memory to allowed nodes.
153 - in vmscan.c, to restrict page recovery to the current cpuset.
154
155You should mount the "cgroup" filesystem type in order to enable
156browsing and modifying the cpusets presently known to the kernel.  No
157new system calls are added for cpusets - all support for querying and
158modifying cpusets is via this cpuset file system.
159
160The /proc/<pid>/status file for each task has four added lines,
161displaying the task's cpus_allowed (on which CPUs it may be scheduled)
162and mems_allowed (on which Memory Nodes it may obtain memory),
163in the two formats seen in the following example::
164
165  Cpus_allowed:   ffffffff,ffffffff,ffffffff,ffffffff
166  Cpus_allowed_list:      0-127
167  Mems_allowed:   ffffffff,ffffffff
168  Mems_allowed_list:      0-63
169
170Each cpuset is represented by a directory in the cgroup file system
171containing (on top of the standard cgroup files) the following
172files describing that cpuset:
173
174 - cpuset.cpus: list of CPUs in that cpuset
175 - cpuset.mems: list of Memory Nodes in that cpuset
176 - cpuset.memory_migrate flag: if set, move pages to cpusets nodes
177 - cpuset.cpu_exclusive flag: is cpu placement exclusive?
178 - cpuset.mem_exclusive flag: is memory placement exclusive?
179 - cpuset.mem_hardwall flag:  is memory allocation hardwalled
180 - cpuset.memory_pressure: measure of how much paging pressure in cpuset
181 - cpuset.memory_spread_page flag: if set, spread page cache evenly on allowed nodes
182 - cpuset.memory_spread_slab flag: if set, spread slab cache evenly on allowed nodes
183 - cpuset.sched_load_balance flag: if set, load balance within CPUs on that cpuset
184 - cpuset.sched_relax_domain_level: the searching range when migrating tasks
185
186In addition, only the root cpuset has the following file:
187
188 - cpuset.memory_pressure_enabled flag: compute memory_pressure?
189
190New cpusets are created using the mkdir system call or shell
191command.  The properties of a cpuset, such as its flags, allowed
192CPUs and Memory Nodes, and attached tasks, are modified by writing
193to the appropriate file in that cpusets directory, as listed above.
194
195The named hierarchical structure of nested cpusets allows partitioning
196a large system into nested, dynamically changeable, "soft-partitions".
197
198The attachment of each task, automatically inherited at fork by any
199children of that task, to a cpuset allows organizing the work load
200on a system into related sets of tasks such that each set is constrained
201to using the CPUs and Memory Nodes of a particular cpuset.  A task
202may be re-attached to any other cpuset, if allowed by the permissions
203on the necessary cpuset file system directories.
204
205Such management of a system "in the large" integrates smoothly with
206the detailed placement done on individual tasks and memory regions
207using the sched_setaffinity, mbind and set_mempolicy system calls.
208
209The following rules apply to each cpuset:
210
211 - Its CPUs and Memory Nodes must be a subset of its parents.
212 - It can't be marked exclusive unless its parent is.
213 - If its cpu or memory is exclusive, they may not overlap any sibling.
214
215These rules, and the natural hierarchy of cpusets, enable efficient
216enforcement of the exclusive guarantee, without having to scan all
217cpusets every time any of them change to ensure nothing overlaps a
218exclusive cpuset.  Also, the use of a Linux virtual file system (vfs)
219to represent the cpuset hierarchy provides for a familiar permission
220and name space for cpusets, with a minimum of additional kernel code.
221
222The cpus and mems files in the root (top_cpuset) cpuset are
223read-only.  The cpus file automatically tracks the value of
224cpu_online_mask using a CPU hotplug notifier, and the mems file
225automatically tracks the value of node_states[N_MEMORY]--i.e.,
226nodes with memory--using the cpuset_track_online_nodes() hook.
227
228The cpuset.effective_cpus and cpuset.effective_mems files are
229normally read-only copies of cpuset.cpus and cpuset.mems files
230respectively.  If the cpuset cgroup filesystem is mounted with the
231special "cpuset_v2_mode" option, the behavior of these files will become
232similar to the corresponding files in cpuset v2.  In other words, hotplug
233events will not change cpuset.cpus and cpuset.mems.  Those events will
234only affect cpuset.effective_cpus and cpuset.effective_mems which show
235the actual cpus and memory nodes that are currently used by this cpuset.
236See Documentation/admin-guide/cgroup-v2.rst for more information about
237cpuset v2 behavior.
238
239
2401.4 What are exclusive cpusets ?
241--------------------------------
242
243If a cpuset is cpu or mem exclusive, no other cpuset, other than
244a direct ancestor or descendant, may share any of the same CPUs or
245Memory Nodes.
246
247A cpuset that is cpuset.mem_exclusive *or* cpuset.mem_hardwall is "hardwalled",
248i.e. it restricts kernel allocations for page, buffer and other data
249commonly shared by the kernel across multiple users.  All cpusets,
250whether hardwalled or not, restrict allocations of memory for user
251space.  This enables configuring a system so that several independent
252jobs can share common kernel data, such as file system pages, while
253isolating each job's user allocation in its own cpuset.  To do this,
254construct a large mem_exclusive cpuset to hold all the jobs, and
255construct child, non-mem_exclusive cpusets for each individual job.
256Only a small amount of typical kernel memory, such as requests from
257interrupt handlers, is allowed to be taken outside even a
258mem_exclusive cpuset.
259
260
2611.5 What is memory_pressure ?
262-----------------------------
263The memory_pressure of a cpuset provides a simple per-cpuset metric
264of the rate that the tasks in a cpuset are attempting to free up in
265use memory on the nodes of the cpuset to satisfy additional memory
266requests.
267
268This enables batch managers monitoring jobs running in dedicated
269cpusets to efficiently detect what level of memory pressure that job
270is causing.
271
272This is useful both on tightly managed systems running a wide mix of
273submitted jobs, which may choose to terminate or re-prioritize jobs that
274are trying to use more memory than allowed on the nodes assigned to them,
275and with tightly coupled, long running, massively parallel scientific
276computing jobs that will dramatically fail to meet required performance
277goals if they start to use more memory than allowed to them.
278
279This mechanism provides a very economical way for the batch manager
280to monitor a cpuset for signs of memory pressure.  It's up to the
281batch manager or other user code to decide what to do about it and
282take action.
283
284==>
285    Unless this feature is enabled by writing "1" to the special file
286    /dev/cpuset/memory_pressure_enabled, the hook in the rebalance
287    code of __alloc_pages() for this metric reduces to simply noticing
288    that the cpuset_memory_pressure_enabled flag is zero.  So only
289    systems that enable this feature will compute the metric.
290
291Why a per-cpuset, running average:
292
293    Because this meter is per-cpuset, rather than per-task or mm,
294    the system load imposed by a batch scheduler monitoring this
295    metric is sharply reduced on large systems, because a scan of
296    the tasklist can be avoided on each set of queries.
297
298    Because this meter is a running average, instead of an accumulating
299    counter, a batch scheduler can detect memory pressure with a
300    single read, instead of having to read and accumulate results
301    for a period of time.
302
303    Because this meter is per-cpuset rather than per-task or mm,
304    the batch scheduler can obtain the key information, memory
305    pressure in a cpuset, with a single read, rather than having to
306    query and accumulate results over all the (dynamically changing)
307    set of tasks in the cpuset.
308
309A per-cpuset simple digital filter (requires a spinlock and 3 words
310of data per-cpuset) is kept, and updated by any task attached to that
311cpuset, if it enters the synchronous (direct) page reclaim code.
312
313A per-cpuset file provides an integer number representing the recent
314(half-life of 10 seconds) rate of direct page reclaims caused by
315the tasks in the cpuset, in units of reclaims attempted per second,
316times 1000.
317
318
3191.6 What is memory spread ?
320---------------------------
321There are two boolean flag files per cpuset that control where the
322kernel allocates pages for the file system buffers and related in
323kernel data structures.  They are called 'cpuset.memory_spread_page' and
324'cpuset.memory_spread_slab'.
325
326If the per-cpuset boolean flag file 'cpuset.memory_spread_page' is set, then
327the kernel will spread the file system buffers (page cache) evenly
328over all the nodes that the faulting task is allowed to use, instead
329of preferring to put those pages on the node where the task is running.
330
331If the per-cpuset boolean flag file 'cpuset.memory_spread_slab' is set,
332then the kernel will spread some file system related slab caches,
333such as for inodes and dentries evenly over all the nodes that the
334faulting task is allowed to use, instead of preferring to put those
335pages on the node where the task is running.
336
337The setting of these flags does not affect anonymous data segment or
338stack segment pages of a task.
339
340By default, both kinds of memory spreading are off, and memory
341pages are allocated on the node local to where the task is running,
342except perhaps as modified by the task's NUMA mempolicy or cpuset
343configuration, so long as sufficient free memory pages are available.
344
345When new cpusets are created, they inherit the memory spread settings
346of their parent.
347
348Setting memory spreading causes allocations for the affected page
349or slab caches to ignore the task's NUMA mempolicy and be spread
350instead.    Tasks using mbind() or set_mempolicy() calls to set NUMA
351mempolicies will not notice any change in these calls as a result of
352their containing task's memory spread settings.  If memory spreading
353is turned off, then the currently specified NUMA mempolicy once again
354applies to memory page allocations.
355
356Both 'cpuset.memory_spread_page' and 'cpuset.memory_spread_slab' are boolean flag
357files.  By default they contain "0", meaning that the feature is off
358for that cpuset.  If a "1" is written to that file, then that turns
359the named feature on.
360
361The implementation is simple.
362
363Setting the flag 'cpuset.memory_spread_page' turns on a per-process flag
364PFA_SPREAD_PAGE for each task that is in that cpuset or subsequently
365joins that cpuset.  The page allocation calls for the page cache
366is modified to perform an inline check for this PFA_SPREAD_PAGE task
367flag, and if set, a call to a new routine cpuset_mem_spread_node()
368returns the node to prefer for the allocation.
369
370Similarly, setting 'cpuset.memory_spread_slab' turns on the flag
371PFA_SPREAD_SLAB, and appropriately marked slab caches will allocate
372pages from the node returned by cpuset_mem_spread_node().
373
374The cpuset_mem_spread_node() routine is also simple.  It uses the
375value of a per-task rotor cpuset_mem_spread_rotor to select the next
376node in the current task's mems_allowed to prefer for the allocation.
377
378This memory placement policy is also known (in other contexts) as
379round-robin or interleave.
380
381This policy can provide substantial improvements for jobs that need
382to place thread local data on the corresponding node, but that need
383to access large file system data sets that need to be spread across
384the several nodes in the jobs cpuset in order to fit.  Without this
385policy, especially for jobs that might have one thread reading in the
386data set, the memory allocation across the nodes in the jobs cpuset
387can become very uneven.
388
3891.7 What is sched_load_balance ?
390--------------------------------
391
392The kernel scheduler (kernel/sched/core.c) automatically load balances
393tasks.  If one CPU is underutilized, kernel code running on that
394CPU will look for tasks on other more overloaded CPUs and move those
395tasks to itself, within the constraints of such placement mechanisms
396as cpusets and sched_setaffinity.
397
398The algorithmic cost of load balancing and its impact on key shared
399kernel data structures such as the task list increases more than
400linearly with the number of CPUs being balanced.  So the scheduler
401has support to partition the systems CPUs into a number of sched
402domains such that it only load balances within each sched domain.
403Each sched domain covers some subset of the CPUs in the system;
404no two sched domains overlap; some CPUs might not be in any sched
405domain and hence won't be load balanced.
406
407Put simply, it costs less to balance between two smaller sched domains
408than one big one, but doing so means that overloads in one of the
409two domains won't be load balanced to the other one.
410
411By default, there is one sched domain covering all CPUs, including those
412marked isolated using the kernel boot time "isolcpus=" argument. However,
413the isolated CPUs will not participate in load balancing, and will not
414have tasks running on them unless explicitly assigned.
415
416This default load balancing across all CPUs is not well suited for
417the following two situations:
418
419 1) On large systems, load balancing across many CPUs is expensive.
420    If the system is managed using cpusets to place independent jobs
421    on separate sets of CPUs, full load balancing is unnecessary.
422 2) Systems supporting realtime on some CPUs need to minimize
423    system overhead on those CPUs, including avoiding task load
424    balancing if that is not needed.
425
426When the per-cpuset flag "cpuset.sched_load_balance" is enabled (the default
427setting), it requests that all the CPUs in that cpusets allowed 'cpuset.cpus'
428be contained in a single sched domain, ensuring that load balancing
429can move a task (not otherwised pinned, as by sched_setaffinity)
430from any CPU in that cpuset to any other.
431
432When the per-cpuset flag "cpuset.sched_load_balance" is disabled, then the
433scheduler will avoid load balancing across the CPUs in that cpuset,
434--except-- in so far as is necessary because some overlapping cpuset
435has "sched_load_balance" enabled.
436
437So, for example, if the top cpuset has the flag "cpuset.sched_load_balance"
438enabled, then the scheduler will have one sched domain covering all
439CPUs, and the setting of the "cpuset.sched_load_balance" flag in any other
440cpusets won't matter, as we're already fully load balancing.
441
442Therefore in the above two situations, the top cpuset flag
443"cpuset.sched_load_balance" should be disabled, and only some of the smaller,
444child cpusets have this flag enabled.
445
446When doing this, you don't usually want to leave any unpinned tasks in
447the top cpuset that might use non-trivial amounts of CPU, as such tasks
448may be artificially constrained to some subset of CPUs, depending on
449the particulars of this flag setting in descendant cpusets.  Even if
450such a task could use spare CPU cycles in some other CPUs, the kernel
451scheduler might not consider the possibility of load balancing that
452task to that underused CPU.
453
454Of course, tasks pinned to a particular CPU can be left in a cpuset
455that disables "cpuset.sched_load_balance" as those tasks aren't going anywhere
456else anyway.
457
458There is an impedance mismatch here, between cpusets and sched domains.
459Cpusets are hierarchical and nest.  Sched domains are flat; they don't
460overlap and each CPU is in at most one sched domain.
461
462It is necessary for sched domains to be flat because load balancing
463across partially overlapping sets of CPUs would risk unstable dynamics
464that would be beyond our understanding.  So if each of two partially
465overlapping cpusets enables the flag 'cpuset.sched_load_balance', then we
466form a single sched domain that is a superset of both.  We won't move
467a task to a CPU outside its cpuset, but the scheduler load balancing
468code might waste some compute cycles considering that possibility.
469
470This mismatch is why there is not a simple one-to-one relation
471between which cpusets have the flag "cpuset.sched_load_balance" enabled,
472and the sched domain configuration.  If a cpuset enables the flag, it
473will get balancing across all its CPUs, but if it disables the flag,
474it will only be assured of no load balancing if no other overlapping
475cpuset enables the flag.
476
477If two cpusets have partially overlapping 'cpuset.cpus' allowed, and only
478one of them has this flag enabled, then the other may find its
479tasks only partially load balanced, just on the overlapping CPUs.
480This is just the general case of the top_cpuset example given a few
481paragraphs above.  In the general case, as in the top cpuset case,
482don't leave tasks that might use non-trivial amounts of CPU in
483such partially load balanced cpusets, as they may be artificially
484constrained to some subset of the CPUs allowed to them, for lack of
485load balancing to the other CPUs.
486
487CPUs in "cpuset.isolcpus" were excluded from load balancing by the
488isolcpus= kernel boot option, and will never be load balanced regardless
489of the value of "cpuset.sched_load_balance" in any cpuset.
490
4911.7.1 sched_load_balance implementation details.
492------------------------------------------------
493
494The per-cpuset flag 'cpuset.sched_load_balance' defaults to enabled (contrary
495to most cpuset flags.)  When enabled for a cpuset, the kernel will
496ensure that it can load balance across all the CPUs in that cpuset
497(makes sure that all the CPUs in the cpus_allowed of that cpuset are
498in the same sched domain.)
499
500If two overlapping cpusets both have 'cpuset.sched_load_balance' enabled,
501then they will be (must be) both in the same sched domain.
502
503If, as is the default, the top cpuset has 'cpuset.sched_load_balance' enabled,
504then by the above that means there is a single sched domain covering
505the whole system, regardless of any other cpuset settings.
506
507The kernel commits to user space that it will avoid load balancing
508where it can.  It will pick as fine a granularity partition of sched
509domains as it can while still providing load balancing for any set
510of CPUs allowed to a cpuset having 'cpuset.sched_load_balance' enabled.
511
512The internal kernel cpuset to scheduler interface passes from the
513cpuset code to the scheduler code a partition of the load balanced
514CPUs in the system. This partition is a set of subsets (represented
515as an array of struct cpumask) of CPUs, pairwise disjoint, that cover
516all the CPUs that must be load balanced.
517
518The cpuset code builds a new such partition and passes it to the
519scheduler sched domain setup code, to have the sched domains rebuilt
520as necessary, whenever:
521
522 - the 'cpuset.sched_load_balance' flag of a cpuset with non-empty CPUs changes,
523 - or CPUs come or go from a cpuset with this flag enabled,
524 - or 'cpuset.sched_relax_domain_level' value of a cpuset with non-empty CPUs
525   and with this flag enabled changes,
526 - or a cpuset with non-empty CPUs and with this flag enabled is removed,
527 - or a cpu is offlined/onlined.
528
529This partition exactly defines what sched domains the scheduler should
530setup - one sched domain for each element (struct cpumask) in the
531partition.
532
533The scheduler remembers the currently active sched domain partitions.
534When the scheduler routine partition_sched_domains() is invoked from
535the cpuset code to update these sched domains, it compares the new
536partition requested with the current, and updates its sched domains,
537removing the old and adding the new, for each change.
538
539
5401.8 What is sched_relax_domain_level ?
541--------------------------------------
542
543In sched domain, the scheduler migrates tasks in 2 ways; periodic load
544balance on tick, and at time of some schedule events.
545
546When a task is woken up, scheduler try to move the task on idle CPU.
547For example, if a task A running on CPU X activates another task B
548on the same CPU X, and if CPU Y is X's sibling and performing idle,
549then scheduler migrate task B to CPU Y so that task B can start on
550CPU Y without waiting task A on CPU X.
551
552And if a CPU run out of tasks in its runqueue, the CPU try to pull
553extra tasks from other busy CPUs to help them before it is going to
554be idle.
555
556Of course it takes some searching cost to find movable tasks and/or
557idle CPUs, the scheduler might not search all CPUs in the domain
558every time.  In fact, in some architectures, the searching ranges on
559events are limited in the same socket or node where the CPU locates,
560while the load balance on tick searches all.
561
562For example, assume CPU Z is relatively far from CPU X.  Even if CPU Z
563is idle while CPU X and the siblings are busy, scheduler can't migrate
564woken task B from X to Z since it is out of its searching range.
565As the result, task B on CPU X need to wait task A or wait load balance
566on the next tick.  For some applications in special situation, waiting
5671 tick may be too long.
568
569The 'cpuset.sched_relax_domain_level' file allows you to request changing
570this searching range as you like.  This file takes int value which
571indicates size of searching range in levels ideally as follows,
572otherwise initial value -1 that indicates the cpuset has no request.
573
574====== ===========================================================
575  -1   no request. use system default or follow request of others.
576   0   no search.
577   1   search siblings (hyperthreads in a core).
578   2   search cores in a package.
579   3   search cpus in a node [= system wide on non-NUMA system]
580   4   search nodes in a chunk of node [on NUMA system]
581   5   search system wide [on NUMA system]
582====== ===========================================================
583
584The system default is architecture dependent.  The system default
585can be changed using the relax_domain_level= boot parameter.
586
587This file is per-cpuset and affect the sched domain where the cpuset
588belongs to.  Therefore if the flag 'cpuset.sched_load_balance' of a cpuset
589is disabled, then 'cpuset.sched_relax_domain_level' have no effect since
590there is no sched domain belonging the cpuset.
591
592If multiple cpusets are overlapping and hence they form a single sched
593domain, the largest value among those is used.  Be careful, if one
594requests 0 and others are -1 then 0 is used.
595
596Note that modifying this file will have both good and bad effects,
597and whether it is acceptable or not depends on your situation.
598Don't modify this file if you are not sure.
599
600If your situation is:
601
602 - The migration costs between each cpu can be assumed considerably
603   small(for you) due to your special application's behavior or
604   special hardware support for CPU cache etc.
605 - The searching cost doesn't have impact(for you) or you can make
606   the searching cost enough small by managing cpuset to compact etc.
607 - The latency is required even it sacrifices cache hit rate etc.
608   then increasing 'sched_relax_domain_level' would benefit you.
609
610
6111.9 How do I use cpusets ?
612--------------------------
613
614In order to minimize the impact of cpusets on critical kernel
615code, such as the scheduler, and due to the fact that the kernel
616does not support one task updating the memory placement of another
617task directly, the impact on a task of changing its cpuset CPU
618or Memory Node placement, or of changing to which cpuset a task
619is attached, is subtle.
620
621If a cpuset has its Memory Nodes modified, then for each task attached
622to that cpuset, the next time that the kernel attempts to allocate
623a page of memory for that task, the kernel will notice the change
624in the task's cpuset, and update its per-task memory placement to
625remain within the new cpusets memory placement.  If the task was using
626mempolicy MPOL_BIND, and the nodes to which it was bound overlap with
627its new cpuset, then the task will continue to use whatever subset
628of MPOL_BIND nodes are still allowed in the new cpuset.  If the task
629was using MPOL_BIND and now none of its MPOL_BIND nodes are allowed
630in the new cpuset, then the task will be essentially treated as if it
631was MPOL_BIND bound to the new cpuset (even though its NUMA placement,
632as queried by get_mempolicy(), doesn't change).  If a task is moved
633from one cpuset to another, then the kernel will adjust the task's
634memory placement, as above, the next time that the kernel attempts
635to allocate a page of memory for that task.
636
637If a cpuset has its 'cpuset.cpus' modified, then each task in that cpuset
638will have its allowed CPU placement changed immediately.  Similarly,
639if a task's pid is written to another cpuset's 'tasks' file, then its
640allowed CPU placement is changed immediately.  If such a task had been
641bound to some subset of its cpuset using the sched_setaffinity() call,
642the task will be allowed to run on any CPU allowed in its new cpuset,
643negating the effect of the prior sched_setaffinity() call.
644
645In summary, the memory placement of a task whose cpuset is changed is
646updated by the kernel, on the next allocation of a page for that task,
647and the processor placement is updated immediately.
648
649Normally, once a page is allocated (given a physical page
650of main memory) then that page stays on whatever node it
651was allocated, so long as it remains allocated, even if the
652cpusets memory placement policy 'cpuset.mems' subsequently changes.
653If the cpuset flag file 'cpuset.memory_migrate' is set true, then when
654tasks are attached to that cpuset, any pages that task had
655allocated to it on nodes in its previous cpuset are migrated
656to the task's new cpuset. The relative placement of the page within
657the cpuset is preserved during these migration operations if possible.
658For example if the page was on the second valid node of the prior cpuset
659then the page will be placed on the second valid node of the new cpuset.
660
661Also if 'cpuset.memory_migrate' is set true, then if that cpuset's
662'cpuset.mems' file is modified, pages allocated to tasks in that
663cpuset, that were on nodes in the previous setting of 'cpuset.mems',
664will be moved to nodes in the new setting of 'mems.'
665Pages that were not in the task's prior cpuset, or in the cpuset's
666prior 'cpuset.mems' setting, will not be moved.
667
668There is an exception to the above.  If hotplug functionality is used
669to remove all the CPUs that are currently assigned to a cpuset,
670then all the tasks in that cpuset will be moved to the nearest ancestor
671with non-empty cpus.  But the moving of some (or all) tasks might fail if
672cpuset is bound with another cgroup subsystem which has some restrictions
673on task attaching.  In this failing case, those tasks will stay
674in the original cpuset, and the kernel will automatically update
675their cpus_allowed to allow all online CPUs.  When memory hotplug
676functionality for removing Memory Nodes is available, a similar exception
677is expected to apply there as well.  In general, the kernel prefers to
678violate cpuset placement, over starving a task that has had all
679its allowed CPUs or Memory Nodes taken offline.
680
681There is a second exception to the above.  GFP_ATOMIC requests are
682kernel internal allocations that must be satisfied, immediately.
683The kernel may drop some request, in rare cases even panic, if a
684GFP_ATOMIC alloc fails.  If the request cannot be satisfied within
685the current task's cpuset, then we relax the cpuset, and look for
686memory anywhere we can find it.  It's better to violate the cpuset
687than stress the kernel.
688
689To start a new job that is to be contained within a cpuset, the steps are:
690
691 1) mkdir /sys/fs/cgroup/cpuset
692 2) mount -t cgroup -ocpuset cpuset /sys/fs/cgroup/cpuset
693 3) Create the new cpuset by doing mkdir's and write's (or echo's) in
694    the /sys/fs/cgroup/cpuset virtual file system.
695 4) Start a task that will be the "founding father" of the new job.
696 5) Attach that task to the new cpuset by writing its pid to the
697    /sys/fs/cgroup/cpuset tasks file for that cpuset.
698 6) fork, exec or clone the job tasks from this founding father task.
699
700For example, the following sequence of commands will setup a cpuset
701named "Charlie", containing just CPUs 2 and 3, and Memory Node 1,
702and then start a subshell 'sh' in that cpuset::
703
704  mount -t cgroup -ocpuset cpuset /sys/fs/cgroup/cpuset
705  cd /sys/fs/cgroup/cpuset
706  mkdir Charlie
707  cd Charlie
708  /bin/echo 2-3 > cpuset.cpus
709  /bin/echo 1 > cpuset.mems
710  /bin/echo $$ > tasks
711  sh
712  # The subshell 'sh' is now running in cpuset Charlie
713  # The next line should display '/Charlie'
714  cat /proc/self/cpuset
715
716There are ways to query or modify cpusets:
717
718 - via the cpuset file system directly, using the various cd, mkdir, echo,
719   cat, rmdir commands from the shell, or their equivalent from C.
720 - via the C library libcpuset.
721 - via the C library libcgroup.
722   (https://github.com/libcgroup/libcgroup/)
723 - via the python application cset.
724   (http://code.google.com/p/cpuset/)
725
726The sched_setaffinity calls can also be done at the shell prompt using
727SGI's runon or Robert Love's taskset.  The mbind and set_mempolicy
728calls can be done at the shell prompt using the numactl command
729(part of Andi Kleen's numa package).
730
7312. Usage Examples and Syntax
732============================
733
7342.1 Basic Usage
735---------------
736
737Creating, modifying, using the cpusets can be done through the cpuset
738virtual filesystem.
739
740To mount it, type:
741# mount -t cgroup -o cpuset cpuset /sys/fs/cgroup/cpuset
742
743Then under /sys/fs/cgroup/cpuset you can find a tree that corresponds to the
744tree of the cpusets in the system. For instance, /sys/fs/cgroup/cpuset
745is the cpuset that holds the whole system.
746
747If you want to create a new cpuset under /sys/fs/cgroup/cpuset::
748
749  # cd /sys/fs/cgroup/cpuset
750  # mkdir my_cpuset
751
752Now you want to do something with this cpuset::
753
754  # cd my_cpuset
755
756In this directory you can find several files::
757
758  # ls
759  cgroup.clone_children  cpuset.memory_pressure
760  cgroup.event_control   cpuset.memory_spread_page
761  cgroup.procs           cpuset.memory_spread_slab
762  cpuset.cpu_exclusive   cpuset.mems
763  cpuset.cpus            cpuset.sched_load_balance
764  cpuset.mem_exclusive   cpuset.sched_relax_domain_level
765  cpuset.mem_hardwall    notify_on_release
766  cpuset.memory_migrate  tasks
767
768Reading them will give you information about the state of this cpuset:
769the CPUs and Memory Nodes it can use, the processes that are using
770it, its properties.  By writing to these files you can manipulate
771the cpuset.
772
773Set some flags::
774
775  # /bin/echo 1 > cpuset.cpu_exclusive
776
777Add some cpus::
778
779  # /bin/echo 0-7 > cpuset.cpus
780
781Add some mems::
782
783  # /bin/echo 0-7 > cpuset.mems
784
785Now attach your shell to this cpuset::
786
787  # /bin/echo $$ > tasks
788
789You can also create cpusets inside your cpuset by using mkdir in this
790directory::
791
792  # mkdir my_sub_cs
793
794To remove a cpuset, just use rmdir::
795
796  # rmdir my_sub_cs
797
798This will fail if the cpuset is in use (has cpusets inside, or has
799processes attached).
800
801Note that for legacy reasons, the "cpuset" filesystem exists as a
802wrapper around the cgroup filesystem.
803
804The command::
805
806  mount -t cpuset X /sys/fs/cgroup/cpuset
807
808is equivalent to::
809
810  mount -t cgroup -ocpuset,noprefix X /sys/fs/cgroup/cpuset
811  echo "/sbin/cpuset_release_agent" > /sys/fs/cgroup/cpuset/release_agent
812
8132.2 Adding/removing cpus
814------------------------
815
816This is the syntax to use when writing in the cpus or mems files
817in cpuset directories::
818
819  # /bin/echo 1-4 > cpuset.cpus		-> set cpus list to cpus 1,2,3,4
820  # /bin/echo 1,2,3,4 > cpuset.cpus	-> set cpus list to cpus 1,2,3,4
821
822To add a CPU to a cpuset, write the new list of CPUs including the
823CPU to be added. To add 6 to the above cpuset::
824
825  # /bin/echo 1-4,6 > cpuset.cpus	-> set cpus list to cpus 1,2,3,4,6
826
827Similarly to remove a CPU from a cpuset, write the new list of CPUs
828without the CPU to be removed.
829
830To remove all the CPUs::
831
832  # /bin/echo "" > cpuset.cpus		-> clear cpus list
833
8342.3 Setting flags
835-----------------
836
837The syntax is very simple::
838
839  # /bin/echo 1 > cpuset.cpu_exclusive 	-> set flag 'cpuset.cpu_exclusive'
840  # /bin/echo 0 > cpuset.cpu_exclusive 	-> unset flag 'cpuset.cpu_exclusive'
841
8422.4 Attaching processes
843-----------------------
844
845::
846
847  # /bin/echo PID > tasks
848
849Note that it is PID, not PIDs. You can only attach ONE task at a time.
850If you have several tasks to attach, you have to do it one after another::
851
852  # /bin/echo PID1 > tasks
853  # /bin/echo PID2 > tasks
854	...
855  # /bin/echo PIDn > tasks
856
857
8583. Questions
859============
860
861Q:
862   what's up with this '/bin/echo' ?
863
864A:
865   bash's builtin 'echo' command does not check calls to write() against
866   errors. If you use it in the cpuset file system, you won't be
867   able to tell whether a command succeeded or failed.
868
869Q:
870   When I attach processes, only the first of the line gets really attached !
871
872A:
873   We can only return one error code per call to write(). So you should also
874   put only ONE pid.
875
8764. Contact
877==========
878
879Web: http://www.bullopensource.org/cpuset
880