blob: aa854b9b18cda8de6fae047540a529b55ca81405 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001 CPUSETS
2 -------
3
4Copyright (C) 2004 BULL SA.
5Written by Simon.Derr@bull.net
6
Christoph Lameterb4fb3762006-03-14 19:50:20 -08007Portions Copyright (c) 2004-2006 Silicon Graphics, Inc.
Linus Torvalds1da177e2005-04-16 15:20:36 -07008Modified by Paul Jackson <pj@sgi.com>
Christoph Lameterb4fb3762006-03-14 19:50:20 -08009Modified by Christoph Lameter <clameter@sgi.com>
Paul Menage8793d852007-10-18 23:39:39 -070010Modified by Paul Menage <menage@google.com>
Hidetoshi Seto4d5f3552008-04-15 14:03:17 +090011Modified by Hidetoshi Seto <seto.hidetoshi@jp.fujitsu.com>
Linus Torvalds1da177e2005-04-16 15:20:36 -070012
13CONTENTS:
14=========
15
161. Cpusets
17 1.1 What are cpusets ?
18 1.2 Why are cpusets needed ?
19 1.3 How are cpusets implemented ?
Paul Jacksonbd5e09c2006-01-08 01:01:50 -080020 1.4 What are exclusive cpusets ?
Paul Menage8793d852007-10-18 23:39:39 -070021 1.5 What is memory_pressure ?
22 1.6 What is memory spread ?
Paul Jackson029190c2007-10-18 23:40:20 -070023 1.7 What is sched_load_balance ?
Hidetoshi Seto4d5f3552008-04-15 14:03:17 +090024 1.8 What is sched_relax_domain_level ?
25 1.9 How do I use cpusets ?
Linus Torvalds1da177e2005-04-16 15:20:36 -0700262. Usage Examples and Syntax
27 2.1 Basic Usage
28 2.2 Adding/removing cpus
29 2.3 Setting flags
30 2.4 Attaching processes
313. Questions
324. Contact
33
341. Cpusets
35==========
36
371.1 What are cpusets ?
38----------------------
39
40Cpusets provide a mechanism for assigning a set of CPUs and Memory
Christoph Lameter0e1e7c72007-10-16 01:25:38 -070041Nodes to a set of tasks. In this document "Memory Node" refers to
42an on-line node that contains memory.
Linus Torvalds1da177e2005-04-16 15:20:36 -070043
44Cpusets constrain the CPU and Memory placement of tasks to only
45the resources within a tasks current cpuset. They form a nested
46hierarchy visible in a virtual file system. These are the essential
47hooks, beyond what is already present, required to manage dynamic
48job placement on large systems.
49
Paul Menage8793d852007-10-18 23:39:39 -070050Cpusets use the generic cgroup subsystem described in
51Documentation/cgroup.txt.
Linus Torvalds1da177e2005-04-16 15:20:36 -070052
Paul Menage8793d852007-10-18 23:39:39 -070053Requests by a task, using the sched_setaffinity(2) system call to
54include CPUs in its CPU affinity mask, and using the mbind(2) and
55set_mempolicy(2) system calls to include Memory Nodes in its memory
56policy, are both filtered through that tasks cpuset, filtering out any
57CPUs or Memory Nodes not in that cpuset. The scheduler will not
58schedule a task on a CPU that is not allowed in its cpus_allowed
59vector, and the kernel page allocator will not allocate a page on a
60node that is not allowed in the requesting tasks mems_allowed vector.
61
62User level code may create and destroy cpusets by name in the cgroup
Linus Torvalds1da177e2005-04-16 15:20:36 -070063virtual file system, manage the attributes and permissions of these
64cpusets and which CPUs and Memory Nodes are assigned to each cpuset,
65specify and query to which cpuset a task is assigned, and list the
66task pids assigned to a cpuset.
67
68
691.2 Why are cpusets needed ?
70----------------------------
71
72The management of large computer systems, with many processors (CPUs),
73complex memory cache hierarchies and multiple Memory Nodes having
74non-uniform access times (NUMA) presents additional challenges for
75the efficient scheduling and memory placement of processes.
76
77Frequently more modest sized systems can be operated with adequate
78efficiency just by letting the operating system automatically share
79the available CPU and Memory resources amongst the requesting tasks.
80
81But larger systems, which benefit more from careful processor and
82memory placement to reduce memory access times and contention,
83and which typically represent a larger investment for the customer,
Jean Delvare33430dc2005-10-30 15:02:20 -080084can benefit from explicitly placing jobs on properly sized subsets of
Linus Torvalds1da177e2005-04-16 15:20:36 -070085the system.
86
87This can be especially valuable on:
88
89 * Web Servers running multiple instances of the same web application,
90 * Servers running different applications (for instance, a web server
91 and a database), or
92 * NUMA systems running large HPC applications with demanding
93 performance characteristics.
94
95These subsets, or "soft partitions" must be able to be dynamically
96adjusted, as the job mix changes, without impacting other concurrently
Christoph Lameterb4fb3762006-03-14 19:50:20 -080097executing jobs. The location of the running jobs pages may also be moved
98when the memory locations are changed.
Linus Torvalds1da177e2005-04-16 15:20:36 -070099
100The kernel cpuset patch provides the minimum essential kernel
101mechanisms required to efficiently implement such subsets. It
102leverages existing CPU and Memory Placement facilities in the Linux
103kernel to avoid any additional impact on the critical scheduler or
104memory allocator code.
105
106
1071.3 How are cpusets implemented ?
108---------------------------------
109
Christoph Lameterb4fb3762006-03-14 19:50:20 -0800110Cpusets provide a Linux kernel mechanism to constrain which CPUs and
111Memory Nodes are used by a process or set of processes.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700112
113The Linux kernel already has a pair of mechanisms to specify on which
114CPUs a task may be scheduled (sched_setaffinity) and on which Memory
115Nodes it may obtain memory (mbind, set_mempolicy).
116
117Cpusets extends these two mechanisms as follows:
118
119 - Cpusets are sets of allowed CPUs and Memory Nodes, known to the
120 kernel.
121 - Each task in the system is attached to a cpuset, via a pointer
Paul Menage8793d852007-10-18 23:39:39 -0700122 in the task structure to a reference counted cgroup structure.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123 - Calls to sched_setaffinity are filtered to just those CPUs
124 allowed in that tasks cpuset.
125 - Calls to mbind and set_mempolicy are filtered to just
126 those Memory Nodes allowed in that tasks cpuset.
127 - The root cpuset contains all the systems CPUs and Memory
128 Nodes.
129 - For any cpuset, one can define child cpusets containing a subset
130 of the parents CPU and Memory Node resources.
131 - The hierarchy of cpusets can be mounted at /dev/cpuset, for
132 browsing and manipulation from user space.
133 - A cpuset may be marked exclusive, which ensures that no other
134 cpuset (except direct ancestors and descendents) may contain
135 any overlapping CPUs or Memory Nodes.
136 - You can list all the tasks (by pid) attached to any cpuset.
137
138The implementation of cpusets requires a few, simple hooks
139into the rest of the kernel, none in performance critical paths:
140
Paul Jackson864913f2006-01-11 02:01:38 +0100141 - in init/main.c, to initialize the root cpuset at system boot.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700142 - in fork and exit, to attach and detach a task from its cpuset.
143 - in sched_setaffinity, to mask the requested CPUs by what's
144 allowed in that tasks cpuset.
145 - in sched.c migrate_all_tasks(), to keep migrating tasks within
146 the CPUs allowed by their cpuset, if possible.
147 - in the mbind and set_mempolicy system calls, to mask the requested
148 Memory Nodes by what's allowed in that tasks cpuset.
Paul Jackson864913f2006-01-11 02:01:38 +0100149 - in page_alloc.c, to restrict memory to allowed nodes.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700150 - in vmscan.c, to restrict page recovery to the current cpuset.
151
Paul Menage8793d852007-10-18 23:39:39 -0700152You should mount the "cgroup" filesystem type in order to enable
153browsing and modifying the cpusets presently known to the kernel. No
154new system calls are added for cpusets - all support for querying and
155modifying cpusets is via this cpuset file system.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156
157The /proc/<pid>/status file for each task has two added lines,
158displaying the tasks cpus_allowed (on which CPUs it may be scheduled)
159and mems_allowed (on which Memory Nodes it may obtain memory),
160in the format seen in the following example:
161
162 Cpus_allowed: ffffffff,ffffffff,ffffffff,ffffffff
163 Mems_allowed: ffffffff,ffffffff
164
Paul Menage8793d852007-10-18 23:39:39 -0700165Each cpuset is represented by a directory in the cgroup file system
166containing (on top of the standard cgroup files) the following
167files describing that cpuset:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700168
169 - cpus: list of CPUs in that cpuset
170 - mems: list of Memory Nodes in that cpuset
Paul Jackson45b07ef2006-01-08 01:00:56 -0800171 - memory_migrate flag: if set, move pages to cpusets nodes
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172 - cpu_exclusive flag: is cpu placement exclusive?
173 - mem_exclusive flag: is memory placement exclusive?
Paul Jacksonbd5e09c2006-01-08 01:01:50 -0800174 - memory_pressure: measure of how much paging pressure in cpuset
175
176In addition, the root cpuset only has the following file:
177 - memory_pressure_enabled flag: compute memory_pressure?
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178
179New cpusets are created using the mkdir system call or shell
180command. The properties of a cpuset, such as its flags, allowed
181CPUs and Memory Nodes, and attached tasks, are modified by writing
182to the appropriate file in that cpusets directory, as listed above.
183
184The named hierarchical structure of nested cpusets allows partitioning
185a large system into nested, dynamically changeable, "soft-partitions".
186
187The attachment of each task, automatically inherited at fork by any
188children of that task, to a cpuset allows organizing the work load
189on a system into related sets of tasks such that each set is constrained
190to using the CPUs and Memory Nodes of a particular cpuset. A task
191may be re-attached to any other cpuset, if allowed by the permissions
192on the necessary cpuset file system directories.
193
194Such management of a system "in the large" integrates smoothly with
195the detailed placement done on individual tasks and memory regions
196using the sched_setaffinity, mbind and set_mempolicy system calls.
197
198The following rules apply to each cpuset:
199
200 - Its CPUs and Memory Nodes must be a subset of its parents.
201 - It can only be marked exclusive if its parent is.
202 - If its cpu or memory is exclusive, they may not overlap any sibling.
203
204These rules, and the natural hierarchy of cpusets, enable efficient
205enforcement of the exclusive guarantee, without having to scan all
206cpusets every time any of them change to ensure nothing overlaps a
207exclusive cpuset. Also, the use of a Linux virtual file system (vfs)
208to represent the cpuset hierarchy provides for a familiar permission
209and name space for cpusets, with a minimum of additional kernel code.
210
Paul Jackson38837fc2006-09-29 02:01:16 -0700211The cpus and mems files in the root (top_cpuset) cpuset are
212read-only. The cpus file automatically tracks the value of
213cpu_online_map using a CPU hotplug notifier, and the mems file
KOSAKI Motohiro0b720372008-02-23 15:23:41 -0800214automatically tracks the value of node_states[N_HIGH_MEMORY]--i.e.,
Christoph Lameter0e1e7c72007-10-16 01:25:38 -0700215nodes with memory--using the cpuset_track_online_nodes() hook.
Paul Jackson4c4d50f2006-08-27 01:23:51 -0700216
Paul Jacksonbd5e09c2006-01-08 01:01:50 -0800217
2181.4 What are exclusive cpusets ?
219--------------------------------
220
221If a cpuset is cpu or mem exclusive, no other cpuset, other than
222a direct ancestor or descendent, may share any of the same CPUs or
223Memory Nodes.
224
Paul Jacksonbd5e09c2006-01-08 01:01:50 -0800225A cpuset that is mem_exclusive restricts kernel allocations for
226page, buffer and other data commonly shared by the kernel across
227multiple users. All cpusets, whether mem_exclusive or not, restrict
228allocations of memory for user space. This enables configuring a
229system so that several independent jobs can share common kernel data,
230such as file system pages, while isolating each jobs user allocation in
231its own cpuset. To do this, construct a large mem_exclusive cpuset to
232hold all the jobs, and construct child, non-mem_exclusive cpusets for
233each individual job. Only a small amount of typical kernel memory,
234such as requests from interrupt handlers, is allowed to be taken
235outside even a mem_exclusive cpuset.
236
237
Paul Menage8793d852007-10-18 23:39:39 -07002381.5 What is memory_pressure ?
Paul Jacksonbd5e09c2006-01-08 01:01:50 -0800239-----------------------------
240The memory_pressure of a cpuset provides a simple per-cpuset metric
241of the rate that the tasks in a cpuset are attempting to free up in
242use memory on the nodes of the cpuset to satisfy additional memory
243requests.
244
245This enables batch managers monitoring jobs running in dedicated
246cpusets to efficiently detect what level of memory pressure that job
247is causing.
248
249This is useful both on tightly managed systems running a wide mix of
250submitted jobs, which may choose to terminate or re-prioritize jobs that
251are trying to use more memory than allowed on the nodes assigned them,
252and with tightly coupled, long running, massively parallel scientific
253computing jobs that will dramatically fail to meet required performance
254goals if they start to use more memory than allowed to them.
255
256This mechanism provides a very economical way for the batch manager
257to monitor a cpuset for signs of memory pressure. It's up to the
258batch manager or other user code to decide what to do about it and
259take action.
260
261==> Unless this feature is enabled by writing "1" to the special file
262 /dev/cpuset/memory_pressure_enabled, the hook in the rebalance
263 code of __alloc_pages() for this metric reduces to simply noticing
264 that the cpuset_memory_pressure_enabled flag is zero. So only
265 systems that enable this feature will compute the metric.
266
267Why a per-cpuset, running average:
268
269 Because this meter is per-cpuset, rather than per-task or mm,
270 the system load imposed by a batch scheduler monitoring this
271 metric is sharply reduced on large systems, because a scan of
272 the tasklist can be avoided on each set of queries.
273
274 Because this meter is a running average, instead of an accumulating
275 counter, a batch scheduler can detect memory pressure with a
276 single read, instead of having to read and accumulate results
277 for a period of time.
278
279 Because this meter is per-cpuset rather than per-task or mm,
280 the batch scheduler can obtain the key information, memory
281 pressure in a cpuset, with a single read, rather than having to
282 query and accumulate results over all the (dynamically changing)
283 set of tasks in the cpuset.
284
285A per-cpuset simple digital filter (requires a spinlock and 3 words
286of data per-cpuset) is kept, and updated by any task attached to that
287cpuset, if it enters the synchronous (direct) page reclaim code.
288
289A per-cpuset file provides an integer number representing the recent
290(half-life of 10 seconds) rate of direct page reclaims caused by
291the tasks in the cpuset, in units of reclaims attempted per second,
292times 1000.
293
294
Paul Menage8793d852007-10-18 23:39:39 -07002951.6 What is memory spread ?
Paul Jackson825a46a2006-03-24 03:16:03 -0800296---------------------------
297There are two boolean flag files per cpuset that control where the
298kernel allocates pages for the file system buffers and related in
299kernel data structures. They are called 'memory_spread_page' and
300'memory_spread_slab'.
301
302If the per-cpuset boolean flag file 'memory_spread_page' is set, then
303the kernel will spread the file system buffers (page cache) evenly
304over all the nodes that the faulting task is allowed to use, instead
305of preferring to put those pages on the node where the task is running.
306
307If the per-cpuset boolean flag file 'memory_spread_slab' is set,
308then the kernel will spread some file system related slab caches,
309such as for inodes and dentries evenly over all the nodes that the
310faulting task is allowed to use, instead of preferring to put those
311pages on the node where the task is running.
312
313The setting of these flags does not affect anonymous data segment or
314stack segment pages of a task.
315
316By default, both kinds of memory spreading are off, and memory
317pages are allocated on the node local to where the task is running,
318except perhaps as modified by the tasks NUMA mempolicy or cpuset
319configuration, so long as sufficient free memory pages are available.
320
321When new cpusets are created, they inherit the memory spread settings
322of their parent.
323
324Setting memory spreading causes allocations for the affected page
325or slab caches to ignore the tasks NUMA mempolicy and be spread
326instead. Tasks using mbind() or set_mempolicy() calls to set NUMA
327mempolicies will not notice any change in these calls as a result of
328their containing tasks memory spread settings. If memory spreading
329is turned off, then the currently specified NUMA mempolicy once again
330applies to memory page allocations.
331
332Both 'memory_spread_page' and 'memory_spread_slab' are boolean flag
333files. By default they contain "0", meaning that the feature is off
334for that cpuset. If a "1" is written to that file, then that turns
335the named feature on.
336
337The implementation is simple.
338
339Setting the flag 'memory_spread_page' turns on a per-process flag
340PF_SPREAD_PAGE for each task that is in that cpuset or subsequently
341joins that cpuset. The page allocation calls for the page cache
342is modified to perform an inline check for this PF_SPREAD_PAGE task
343flag, and if set, a call to a new routine cpuset_mem_spread_node()
344returns the node to prefer for the allocation.
345
346Similarly, setting 'memory_spread_cache' turns on the flag
347PF_SPREAD_SLAB, and appropriately marked slab caches will allocate
348pages from the node returned by cpuset_mem_spread_node().
349
350The cpuset_mem_spread_node() routine is also simple. It uses the
351value of a per-task rotor cpuset_mem_spread_rotor to select the next
352node in the current tasks mems_allowed to prefer for the allocation.
353
354This memory placement policy is also known (in other contexts) as
355round-robin or interleave.
356
357This policy can provide substantial improvements for jobs that need
358to place thread local data on the corresponding node, but that need
359to access large file system data sets that need to be spread across
360the several nodes in the jobs cpuset in order to fit. Without this
361policy, especially for jobs that might have one thread reading in the
362data set, the memory allocation across the nodes in the jobs cpuset
363can become very uneven.
364
Paul Jackson029190c2007-10-18 23:40:20 -07003651.7 What is sched_load_balance ?
366--------------------------------
Paul Jackson825a46a2006-03-24 03:16:03 -0800367
Paul Jackson029190c2007-10-18 23:40:20 -0700368The kernel scheduler (kernel/sched.c) automatically load balances
369tasks. If one CPU is underutilized, kernel code running on that
370CPU will look for tasks on other more overloaded CPUs and move those
371tasks to itself, within the constraints of such placement mechanisms
372as cpusets and sched_setaffinity.
373
374The algorithmic cost of load balancing and its impact on key shared
375kernel data structures such as the task list increases more than
376linearly with the number of CPUs being balanced. So the scheduler
377has support to partition the systems CPUs into a number of sched
378domains such that it only load balances within each sched domain.
379Each sched domain covers some subset of the CPUs in the system;
380no two sched domains overlap; some CPUs might not be in any sched
381domain and hence won't be load balanced.
382
383Put simply, it costs less to balance between two smaller sched domains
384than one big one, but doing so means that overloads in one of the
385two domains won't be load balanced to the other one.
386
387By default, there is one sched domain covering all CPUs, except those
388marked isolated using the kernel boot time "isolcpus=" argument.
389
390This default load balancing across all CPUs is not well suited for
391the following two situations:
392 1) On large systems, load balancing across many CPUs is expensive.
393 If the system is managed using cpusets to place independent jobs
394 on separate sets of CPUs, full load balancing is unnecessary.
395 2) Systems supporting realtime on some CPUs need to minimize
396 system overhead on those CPUs, including avoiding task load
397 balancing if that is not needed.
398
399When the per-cpuset flag "sched_load_balance" is enabled (the default
400setting), it requests that all the CPUs in that cpusets allowed 'cpus'
401be contained in a single sched domain, ensuring that load balancing
402can move a task (not otherwised pinned, as by sched_setaffinity)
403from any CPU in that cpuset to any other.
404
405When the per-cpuset flag "sched_load_balance" is disabled, then the
406scheduler will avoid load balancing across the CPUs in that cpuset,
407--except-- in so far as is necessary because some overlapping cpuset
408has "sched_load_balance" enabled.
409
410So, for example, if the top cpuset has the flag "sched_load_balance"
411enabled, then the scheduler will have one sched domain covering all
412CPUs, and the setting of the "sched_load_balance" flag in any other
413cpusets won't matter, as we're already fully load balancing.
414
415Therefore in the above two situations, the top cpuset flag
416"sched_load_balance" should be disabled, and only some of the smaller,
417child cpusets have this flag enabled.
418
419When doing this, you don't usually want to leave any unpinned tasks in
420the top cpuset that might use non-trivial amounts of CPU, as such tasks
421may be artificially constrained to some subset of CPUs, depending on
422the particulars of this flag setting in descendent cpusets. Even if
423such a task could use spare CPU cycles in some other CPUs, the kernel
424scheduler might not consider the possibility of load balancing that
425task to that underused CPU.
426
427Of course, tasks pinned to a particular CPU can be left in a cpuset
428that disables "sched_load_balance" as those tasks aren't going anywhere
429else anyway.
430
431There is an impedance mismatch here, between cpusets and sched domains.
432Cpusets are hierarchical and nest. Sched domains are flat; they don't
433overlap and each CPU is in at most one sched domain.
434
435It is necessary for sched domains to be flat because load balancing
436across partially overlapping sets of CPUs would risk unstable dynamics
437that would be beyond our understanding. So if each of two partially
438overlapping cpusets enables the flag 'sched_load_balance', then we
439form a single sched domain that is a superset of both. We won't move
440a task to a CPU outside it cpuset, but the scheduler load balancing
441code might waste some compute cycles considering that possibility.
442
443This mismatch is why there is not a simple one-to-one relation
444between which cpusets have the flag "sched_load_balance" enabled,
445and the sched domain configuration. If a cpuset enables the flag, it
446will get balancing across all its CPUs, but if it disables the flag,
447it will only be assured of no load balancing if no other overlapping
448cpuset enables the flag.
449
450If two cpusets have partially overlapping 'cpus' allowed, and only
451one of them has this flag enabled, then the other may find its
452tasks only partially load balanced, just on the overlapping CPUs.
453This is just the general case of the top_cpuset example given a few
454paragraphs above. In the general case, as in the top cpuset case,
455don't leave tasks that might use non-trivial amounts of CPU in
456such partially load balanced cpusets, as they may be artificially
457constrained to some subset of the CPUs allowed to them, for lack of
458load balancing to the other CPUs.
459
4601.7.1 sched_load_balance implementation details.
461------------------------------------------------
462
463The per-cpuset flag 'sched_load_balance' defaults to enabled (contrary
464to most cpuset flags.) When enabled for a cpuset, the kernel will
465ensure that it can load balance across all the CPUs in that cpuset
466(makes sure that all the CPUs in the cpus_allowed of that cpuset are
467in the same sched domain.)
468
469If two overlapping cpusets both have 'sched_load_balance' enabled,
470then they will be (must be) both in the same sched domain.
471
472If, as is the default, the top cpuset has 'sched_load_balance' enabled,
473then by the above that means there is a single sched domain covering
474the whole system, regardless of any other cpuset settings.
475
476The kernel commits to user space that it will avoid load balancing
477where it can. It will pick as fine a granularity partition of sched
478domains as it can while still providing load balancing for any set
479of CPUs allowed to a cpuset having 'sched_load_balance' enabled.
480
481The internal kernel cpuset to scheduler interface passes from the
482cpuset code to the scheduler code a partition of the load balanced
483CPUs in the system. This partition is a set of subsets (represented
484as an array of cpumask_t) of CPUs, pairwise disjoint, that cover all
485the CPUs that must be load balanced.
486
487Whenever the 'sched_load_balance' flag changes, or CPUs come or go
488from a cpuset with this flag enabled, or a cpuset with this flag
489enabled is removed, the cpuset code builds a new such partition and
490passes it to the scheduler sched domain setup code, to have the sched
491domains rebuilt as necessary.
492
493This partition exactly defines what sched domains the scheduler should
494setup - one sched domain for each element (cpumask_t) in the partition.
495
496The scheduler remembers the currently active sched domain partitions.
497When the scheduler routine partition_sched_domains() is invoked from
498the cpuset code to update these sched domains, it compares the new
499partition requested with the current, and updates its sched domains,
500removing the old and adding the new, for each change.
501
Hidetoshi Seto4d5f3552008-04-15 14:03:17 +0900502
5031.8 What is sched_relax_domain_level ?
504--------------------------------------
505
506In sched domain, the scheduler migrates tasks in 2 ways; periodic load
507balance on tick, and at time of some schedule events.
508
509When a task is woken up, scheduler try to move the task on idle CPU.
510For example, if a task A running on CPU X activates another task B
511on the same CPU X, and if CPU Y is X's sibling and performing idle,
512then scheduler migrate task B to CPU Y so that task B can start on
513CPU Y without waiting task A on CPU X.
514
515And if a CPU run out of tasks in its runqueue, the CPU try to pull
516extra tasks from other busy CPUs to help them before it is going to
517be idle.
518
519Of course it takes some searching cost to find movable tasks and/or
520idle CPUs, the scheduler might not search all CPUs in the domain
521everytime. In fact, in some architectures, the searching ranges on
522events are limited in the same socket or node where the CPU locates,
523while the load balance on tick searchs all.
524
525For example, assume CPU Z is relatively far from CPU X. Even if CPU Z
526is idle while CPU X and the siblings are busy, scheduler can't migrate
527woken task B from X to Z since it is out of its searching range.
528As the result, task B on CPU X need to wait task A or wait load balance
529on the next tick. For some applications in special situation, waiting
5301 tick may be too long.
531
532The 'sched_relax_domain_level' file allows you to request changing
533this searching range as you like. This file takes int value which
534indicates size of searching range in levels ideally as follows,
535otherwise initial value -1 that indicates the cpuset has no request.
536
537 -1 : no request. use system default or follow request of others.
538 0 : no search.
539 1 : search siblings (hyperthreads in a core).
540 2 : search cores in a package.
541 3 : search cpus in a node [= system wide on non-NUMA system]
542 ( 4 : search nodes in a chunk of node [on NUMA system] )
543 ( 5~ : search system wide [on NUMA system])
544
545This file is per-cpuset and affect the sched domain where the cpuset
546belongs to. Therefore if the flag 'sched_load_balance' of a cpuset
547is disabled, then 'sched_relax_domain_level' have no effect since
548there is no sched domain belonging the cpuset.
549
550If multiple cpusets are overlapping and hence they form a single sched
551domain, the largest value among those is used. Be careful, if one
552requests 0 and others are -1 then 0 is used.
553
554Note that modifying this file will have both good and bad effects,
555and whether it is acceptable or not will be depend on your situation.
556Don't modify this file if you are not sure.
557
558If your situation is:
559 - The migration costs between each cpu can be assumed considerably
560 small(for you) due to your special application's behavior or
561 special hardware support for CPU cache etc.
562 - The searching cost doesn't have impact(for you) or you can make
563 the searching cost enough small by managing cpuset to compact etc.
564 - The latency is required even it sacrifices cache hit rate etc.
565then increasing 'sched_relax_domain_level' would benefit you.
566
567
5681.9 How do I use cpusets ?
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569--------------------------
570
571In order to minimize the impact of cpusets on critical kernel
572code, such as the scheduler, and due to the fact that the kernel
573does not support one task updating the memory placement of another
574task directly, the impact on a task of changing its cpuset CPU
575or Memory Node placement, or of changing to which cpuset a task
576is attached, is subtle.
577
578If a cpuset has its Memory Nodes modified, then for each task attached
579to that cpuset, the next time that the kernel attempts to allocate
580a page of memory for that task, the kernel will notice the change
581in the tasks cpuset, and update its per-task memory placement to
582remain within the new cpusets memory placement. If the task was using
583mempolicy MPOL_BIND, and the nodes to which it was bound overlap with
584its new cpuset, then the task will continue to use whatever subset
585of MPOL_BIND nodes are still allowed in the new cpuset. If the task
586was using MPOL_BIND and now none of its MPOL_BIND nodes are allowed
587in the new cpuset, then the task will be essentially treated as if it
588was MPOL_BIND bound to the new cpuset (even though its numa placement,
589as queried by get_mempolicy(), doesn't change). If a task is moved
590from one cpuset to another, then the kernel will adjust the tasks
591memory placement, as above, the next time that the kernel attempts
592to allocate a page of memory for that task.
593
Paul Jackson8f5aa262008-02-07 00:14:48 -0800594If a cpuset has its 'cpus' modified, then each task in that cpuset
595will have its allowed CPU placement changed immediately. Similarly,
596if a tasks pid is written to a cpusets 'tasks' file, in either its
597current cpuset or another cpuset, then its allowed CPU placement is
598changed immediately. If such a task had been bound to some subset
599of its cpuset using the sched_setaffinity() call, the task will be
600allowed to run on any CPU allowed in its new cpuset, negating the
601affect of the prior sched_setaffinity() call.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700602
603In summary, the memory placement of a task whose cpuset is changed is
604updated by the kernel, on the next allocation of a page for that task,
605but the processor placement is not updated, until that tasks pid is
606rewritten to the 'tasks' file of its cpuset. This is done to avoid
607impacting the scheduler code in the kernel with a check for changes
608in a tasks processor placement.
609
Paul Jackson45b07ef2006-01-08 01:00:56 -0800610Normally, once a page is allocated (given a physical page
611of main memory) then that page stays on whatever node it
612was allocated, so long as it remains allocated, even if the
613cpusets memory placement policy 'mems' subsequently changes.
614If the cpuset flag file 'memory_migrate' is set true, then when
615tasks are attached to that cpuset, any pages that task had
616allocated to it on nodes in its previous cpuset are migrated
Christoph Lameterb4fb3762006-03-14 19:50:20 -0800617to the tasks new cpuset. The relative placement of the page within
618the cpuset is preserved during these migration operations if possible.
619For example if the page was on the second valid node of the prior cpuset
620then the page will be placed on the second valid node of the new cpuset.
621
Paul Jackson45b07ef2006-01-08 01:00:56 -0800622Also if 'memory_migrate' is set true, then if that cpusets
623'mems' file is modified, pages allocated to tasks in that
624cpuset, that were on nodes in the previous setting of 'mems',
Christoph Lameterb4fb3762006-03-14 19:50:20 -0800625will be moved to nodes in the new setting of 'mems.'
626Pages that were not in the tasks prior cpuset, or in the cpusets
627prior 'mems' setting, will not be moved.
Paul Jackson45b07ef2006-01-08 01:00:56 -0800628
Tobias Klauserd533f672005-09-10 00:26:46 -0700629There is an exception to the above. If hotplug functionality is used
Linus Torvalds1da177e2005-04-16 15:20:36 -0700630to remove all the CPUs that are currently assigned to a cpuset,
631then the kernel will automatically update the cpus_allowed of all
Paul Jacksonb39c4fa2005-05-20 13:59:15 -0700632tasks attached to CPUs in that cpuset to allow all CPUs. When memory
Linus Torvalds1da177e2005-04-16 15:20:36 -0700633hotplug functionality for removing Memory Nodes is available, a
634similar exception is expected to apply there as well. In general,
635the kernel prefers to violate cpuset placement, over starving a task
636that has had all its allowed CPUs or Memory Nodes taken offline. User
637code should reconfigure cpusets to only refer to online CPUs and Memory
638Nodes when using hotplug to add or remove such resources.
639
640There is a second exception to the above. GFP_ATOMIC requests are
641kernel internal allocations that must be satisfied, immediately.
642The kernel may drop some request, in rare cases even panic, if a
643GFP_ATOMIC alloc fails. If the request cannot be satisfied within
644the current tasks cpuset, then we relax the cpuset, and look for
645memory anywhere we can find it. It's better to violate the cpuset
646than stress the kernel.
647
648To start a new job that is to be contained within a cpuset, the steps are:
649
650 1) mkdir /dev/cpuset
Paul Menage8793d852007-10-18 23:39:39 -0700651 2) mount -t cgroup -ocpuset cpuset /dev/cpuset
Linus Torvalds1da177e2005-04-16 15:20:36 -0700652 3) Create the new cpuset by doing mkdir's and write's (or echo's) in
653 the /dev/cpuset virtual file system.
654 4) Start a task that will be the "founding father" of the new job.
655 5) Attach that task to the new cpuset by writing its pid to the
656 /dev/cpuset tasks file for that cpuset.
657 6) fork, exec or clone the job tasks from this founding father task.
658
659For example, the following sequence of commands will setup a cpuset
660named "Charlie", containing just CPUs 2 and 3, and Memory Node 1,
661and then start a subshell 'sh' in that cpuset:
662
Paul Menage8793d852007-10-18 23:39:39 -0700663 mount -t cgroup -ocpuset cpuset /dev/cpuset
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664 cd /dev/cpuset
665 mkdir Charlie
666 cd Charlie
667 /bin/echo 2-3 > cpus
668 /bin/echo 1 > mems
669 /bin/echo $$ > tasks
670 sh
671 # The subshell 'sh' is now running in cpuset Charlie
672 # The next line should display '/Charlie'
673 cat /proc/self/cpuset
674
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675In the future, a C library interface to cpusets will likely be
676available. For now, the only way to query or modify cpusets is
677via the cpuset file system, using the various cd, mkdir, echo, cat,
678rmdir commands from the shell, or their equivalent from C.
679
680The sched_setaffinity calls can also be done at the shell prompt using
681SGI's runon or Robert Love's taskset. The mbind and set_mempolicy
682calls can be done at the shell prompt using the numactl command
683(part of Andi Kleen's numa package).
684
6852. Usage Examples and Syntax
686============================
687
6882.1 Basic Usage
689---------------
690
691Creating, modifying, using the cpusets can be done through the cpuset
692virtual filesystem.
693
694To mount it, type:
Paul Menage8793d852007-10-18 23:39:39 -0700695# mount -t cgroup -o cpuset cpuset /dev/cpuset
Linus Torvalds1da177e2005-04-16 15:20:36 -0700696
697Then under /dev/cpuset you can find a tree that corresponds to the
698tree of the cpusets in the system. For instance, /dev/cpuset
699is the cpuset that holds the whole system.
700
701If you want to create a new cpuset under /dev/cpuset:
702# cd /dev/cpuset
703# mkdir my_cpuset
704
705Now you want to do something with this cpuset.
706# cd my_cpuset
707
708In this directory you can find several files:
709# ls
710cpus cpu_exclusive mems mem_exclusive tasks
711
712Reading them will give you information about the state of this cpuset:
713the CPUs and Memory Nodes it can use, the processes that are using
714it, its properties. By writing to these files you can manipulate
715the cpuset.
716
717Set some flags:
718# /bin/echo 1 > cpu_exclusive
719
720Add some cpus:
721# /bin/echo 0-7 > cpus
722
Simon Horman2400ff72007-04-01 23:49:40 -0700723Add some mems:
724# /bin/echo 0-7 > mems
725
Linus Torvalds1da177e2005-04-16 15:20:36 -0700726Now attach your shell to this cpuset:
727# /bin/echo $$ > tasks
728
729You can also create cpusets inside your cpuset by using mkdir in this
730directory.
731# mkdir my_sub_cs
732
733To remove a cpuset, just use rmdir:
734# rmdir my_sub_cs
735This will fail if the cpuset is in use (has cpusets inside, or has
736processes attached).
737
Paul Menage8793d852007-10-18 23:39:39 -0700738Note that for legacy reasons, the "cpuset" filesystem exists as a
739wrapper around the cgroup filesystem.
740
741The command
742
743mount -t cpuset X /dev/cpuset
744
745is equivalent to
746
747mount -t cgroup -ocpuset X /dev/cpuset
748echo "/sbin/cpuset_release_agent" > /dev/cpuset/release_agent
749
Linus Torvalds1da177e2005-04-16 15:20:36 -07007502.2 Adding/removing cpus
751------------------------
752
753This is the syntax to use when writing in the cpus or mems files
754in cpuset directories:
755
756# /bin/echo 1-4 > cpus -> set cpus list to cpus 1,2,3,4
757# /bin/echo 1,2,3,4 > cpus -> set cpus list to cpus 1,2,3,4
758
7592.3 Setting flags
760-----------------
761
762The syntax is very simple:
763
764# /bin/echo 1 > cpu_exclusive -> set flag 'cpu_exclusive'
765# /bin/echo 0 > cpu_exclusive -> unset flag 'cpu_exclusive'
766
7672.4 Attaching processes
768-----------------------
769
770# /bin/echo PID > tasks
771
772Note that it is PID, not PIDs. You can only attach ONE task at a time.
773If you have several tasks to attach, you have to do it one after another:
774
775# /bin/echo PID1 > tasks
776# /bin/echo PID2 > tasks
777 ...
778# /bin/echo PIDn > tasks
779
780
7813. Questions
782============
783
784Q: what's up with this '/bin/echo' ?
785A: bash's builtin 'echo' command does not check calls to write() against
786 errors. If you use it in the cpuset file system, you won't be
787 able to tell whether a command succeeded or failed.
788
789Q: When I attach processes, only the first of the line gets really attached !
790A: We can only return one error code per call to write(). So you should also
791 put only ONE pid.
792
7934. Contact
794==========
795
796Web: http://www.bullopensource.org/cpuset