blob: fb7b361e6eea66c573af07d86c415ad4371dcb64 [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 Menage78608362008-04-29 01:00:26 -0700174 - mem_hardwall flag: is memory allocation hardwalled
Paul Jacksonbd5e09c2006-01-08 01:01:50 -0800175 - memory_pressure: measure of how much paging pressure in cpuset
176
177In addition, the root cpuset only has the following file:
178 - memory_pressure_enabled flag: compute memory_pressure?
Linus Torvalds1da177e2005-04-16 15:20:36 -0700179
180New cpusets are created using the mkdir system call or shell
181command. The properties of a cpuset, such as its flags, allowed
182CPUs and Memory Nodes, and attached tasks, are modified by writing
183to the appropriate file in that cpusets directory, as listed above.
184
185The named hierarchical structure of nested cpusets allows partitioning
186a large system into nested, dynamically changeable, "soft-partitions".
187
188The attachment of each task, automatically inherited at fork by any
189children of that task, to a cpuset allows organizing the work load
190on a system into related sets of tasks such that each set is constrained
191to using the CPUs and Memory Nodes of a particular cpuset. A task
192may be re-attached to any other cpuset, if allowed by the permissions
193on the necessary cpuset file system directories.
194
195Such management of a system "in the large" integrates smoothly with
196the detailed placement done on individual tasks and memory regions
197using the sched_setaffinity, mbind and set_mempolicy system calls.
198
199The following rules apply to each cpuset:
200
201 - Its CPUs and Memory Nodes must be a subset of its parents.
202 - It can only be marked exclusive if its parent is.
203 - If its cpu or memory is exclusive, they may not overlap any sibling.
204
205These rules, and the natural hierarchy of cpusets, enable efficient
206enforcement of the exclusive guarantee, without having to scan all
207cpusets every time any of them change to ensure nothing overlaps a
208exclusive cpuset. Also, the use of a Linux virtual file system (vfs)
209to represent the cpuset hierarchy provides for a familiar permission
210and name space for cpusets, with a minimum of additional kernel code.
211
Paul Jackson38837fc2006-09-29 02:01:16 -0700212The cpus and mems files in the root (top_cpuset) cpuset are
213read-only. The cpus file automatically tracks the value of
214cpu_online_map using a CPU hotplug notifier, and the mems file
KOSAKI Motohiro0b720372008-02-23 15:23:41 -0800215automatically tracks the value of node_states[N_HIGH_MEMORY]--i.e.,
Christoph Lameter0e1e7c72007-10-16 01:25:38 -0700216nodes with memory--using the cpuset_track_online_nodes() hook.
Paul Jackson4c4d50f2006-08-27 01:23:51 -0700217
Paul Jacksonbd5e09c2006-01-08 01:01:50 -0800218
2191.4 What are exclusive cpusets ?
220--------------------------------
221
222If a cpuset is cpu or mem exclusive, no other cpuset, other than
223a direct ancestor or descendent, may share any of the same CPUs or
224Memory Nodes.
225
Paul Menage78608362008-04-29 01:00:26 -0700226A cpuset that is mem_exclusive *or* mem_hardwall is "hardwalled",
227i.e. it restricts kernel allocations for page, buffer and other data
228commonly shared by the kernel across multiple users. All cpusets,
229whether hardwalled or not, restrict allocations of memory for user
230space. This enables configuring a system so that several independent
231jobs can share common kernel data, such as file system pages, while
232isolating each job's user allocation in its own cpuset. To do this,
233construct a large mem_exclusive cpuset to hold all the jobs, and
234construct child, non-mem_exclusive cpusets for each individual job.
235Only a small amount of typical kernel memory, such as requests from
236interrupt handlers, is allowed to be taken outside even a
237mem_exclusive cpuset.
Paul Jacksonbd5e09c2006-01-08 01:01:50 -0800238
239
Paul Menage8793d852007-10-18 23:39:39 -07002401.5 What is memory_pressure ?
Paul Jacksonbd5e09c2006-01-08 01:01:50 -0800241-----------------------------
242The memory_pressure of a cpuset provides a simple per-cpuset metric
243of the rate that the tasks in a cpuset are attempting to free up in
244use memory on the nodes of the cpuset to satisfy additional memory
245requests.
246
247This enables batch managers monitoring jobs running in dedicated
248cpusets to efficiently detect what level of memory pressure that job
249is causing.
250
251This is useful both on tightly managed systems running a wide mix of
252submitted jobs, which may choose to terminate or re-prioritize jobs that
253are trying to use more memory than allowed on the nodes assigned them,
254and with tightly coupled, long running, massively parallel scientific
255computing jobs that will dramatically fail to meet required performance
256goals if they start to use more memory than allowed to them.
257
258This mechanism provides a very economical way for the batch manager
259to monitor a cpuset for signs of memory pressure. It's up to the
260batch manager or other user code to decide what to do about it and
261take action.
262
263==> Unless this feature is enabled by writing "1" to the special file
264 /dev/cpuset/memory_pressure_enabled, the hook in the rebalance
265 code of __alloc_pages() for this metric reduces to simply noticing
266 that the cpuset_memory_pressure_enabled flag is zero. So only
267 systems that enable this feature will compute the metric.
268
269Why a per-cpuset, running average:
270
271 Because this meter is per-cpuset, rather than per-task or mm,
272 the system load imposed by a batch scheduler monitoring this
273 metric is sharply reduced on large systems, because a scan of
274 the tasklist can be avoided on each set of queries.
275
276 Because this meter is a running average, instead of an accumulating
277 counter, a batch scheduler can detect memory pressure with a
278 single read, instead of having to read and accumulate results
279 for a period of time.
280
281 Because this meter is per-cpuset rather than per-task or mm,
282 the batch scheduler can obtain the key information, memory
283 pressure in a cpuset, with a single read, rather than having to
284 query and accumulate results over all the (dynamically changing)
285 set of tasks in the cpuset.
286
287A per-cpuset simple digital filter (requires a spinlock and 3 words
288of data per-cpuset) is kept, and updated by any task attached to that
289cpuset, if it enters the synchronous (direct) page reclaim code.
290
291A per-cpuset file provides an integer number representing the recent
292(half-life of 10 seconds) rate of direct page reclaims caused by
293the tasks in the cpuset, in units of reclaims attempted per second,
294times 1000.
295
296
Paul Menage8793d852007-10-18 23:39:39 -07002971.6 What is memory spread ?
Paul Jackson825a46a2006-03-24 03:16:03 -0800298---------------------------
299There are two boolean flag files per cpuset that control where the
300kernel allocates pages for the file system buffers and related in
301kernel data structures. They are called 'memory_spread_page' and
302'memory_spread_slab'.
303
304If the per-cpuset boolean flag file 'memory_spread_page' is set, then
305the kernel will spread the file system buffers (page cache) evenly
306over all the nodes that the faulting task is allowed to use, instead
307of preferring to put those pages on the node where the task is running.
308
309If the per-cpuset boolean flag file 'memory_spread_slab' is set,
310then the kernel will spread some file system related slab caches,
311such as for inodes and dentries evenly over all the nodes that the
312faulting task is allowed to use, instead of preferring to put those
313pages on the node where the task is running.
314
315The setting of these flags does not affect anonymous data segment or
316stack segment pages of a task.
317
318By default, both kinds of memory spreading are off, and memory
319pages are allocated on the node local to where the task is running,
320except perhaps as modified by the tasks NUMA mempolicy or cpuset
321configuration, so long as sufficient free memory pages are available.
322
323When new cpusets are created, they inherit the memory spread settings
324of their parent.
325
326Setting memory spreading causes allocations for the affected page
327or slab caches to ignore the tasks NUMA mempolicy and be spread
328instead. Tasks using mbind() or set_mempolicy() calls to set NUMA
329mempolicies will not notice any change in these calls as a result of
330their containing tasks memory spread settings. If memory spreading
331is turned off, then the currently specified NUMA mempolicy once again
332applies to memory page allocations.
333
334Both 'memory_spread_page' and 'memory_spread_slab' are boolean flag
335files. By default they contain "0", meaning that the feature is off
336for that cpuset. If a "1" is written to that file, then that turns
337the named feature on.
338
339The implementation is simple.
340
341Setting the flag 'memory_spread_page' turns on a per-process flag
342PF_SPREAD_PAGE for each task that is in that cpuset or subsequently
343joins that cpuset. The page allocation calls for the page cache
344is modified to perform an inline check for this PF_SPREAD_PAGE task
345flag, and if set, a call to a new routine cpuset_mem_spread_node()
346returns the node to prefer for the allocation.
347
348Similarly, setting 'memory_spread_cache' turns on the flag
349PF_SPREAD_SLAB, and appropriately marked slab caches will allocate
350pages from the node returned by cpuset_mem_spread_node().
351
352The cpuset_mem_spread_node() routine is also simple. It uses the
353value of a per-task rotor cpuset_mem_spread_rotor to select the next
354node in the current tasks mems_allowed to prefer for the allocation.
355
356This memory placement policy is also known (in other contexts) as
357round-robin or interleave.
358
359This policy can provide substantial improvements for jobs that need
360to place thread local data on the corresponding node, but that need
361to access large file system data sets that need to be spread across
362the several nodes in the jobs cpuset in order to fit. Without this
363policy, especially for jobs that might have one thread reading in the
364data set, the memory allocation across the nodes in the jobs cpuset
365can become very uneven.
366
Paul Jackson029190c2007-10-18 23:40:20 -07003671.7 What is sched_load_balance ?
368--------------------------------
Paul Jackson825a46a2006-03-24 03:16:03 -0800369
Paul Jackson029190c2007-10-18 23:40:20 -0700370The kernel scheduler (kernel/sched.c) automatically load balances
371tasks. If one CPU is underutilized, kernel code running on that
372CPU will look for tasks on other more overloaded CPUs and move those
373tasks to itself, within the constraints of such placement mechanisms
374as cpusets and sched_setaffinity.
375
376The algorithmic cost of load balancing and its impact on key shared
377kernel data structures such as the task list increases more than
378linearly with the number of CPUs being balanced. So the scheduler
379has support to partition the systems CPUs into a number of sched
380domains such that it only load balances within each sched domain.
381Each sched domain covers some subset of the CPUs in the system;
382no two sched domains overlap; some CPUs might not be in any sched
383domain and hence won't be load balanced.
384
385Put simply, it costs less to balance between two smaller sched domains
386than one big one, but doing so means that overloads in one of the
387two domains won't be load balanced to the other one.
388
389By default, there is one sched domain covering all CPUs, except those
390marked isolated using the kernel boot time "isolcpus=" argument.
391
392This default load balancing across all CPUs is not well suited for
393the following two situations:
394 1) On large systems, load balancing across many CPUs is expensive.
395 If the system is managed using cpusets to place independent jobs
396 on separate sets of CPUs, full load balancing is unnecessary.
397 2) Systems supporting realtime on some CPUs need to minimize
398 system overhead on those CPUs, including avoiding task load
399 balancing if that is not needed.
400
401When the per-cpuset flag "sched_load_balance" is enabled (the default
402setting), it requests that all the CPUs in that cpusets allowed 'cpus'
403be contained in a single sched domain, ensuring that load balancing
404can move a task (not otherwised pinned, as by sched_setaffinity)
405from any CPU in that cpuset to any other.
406
407When the per-cpuset flag "sched_load_balance" is disabled, then the
408scheduler will avoid load balancing across the CPUs in that cpuset,
409--except-- in so far as is necessary because some overlapping cpuset
410has "sched_load_balance" enabled.
411
412So, for example, if the top cpuset has the flag "sched_load_balance"
413enabled, then the scheduler will have one sched domain covering all
414CPUs, and the setting of the "sched_load_balance" flag in any other
415cpusets won't matter, as we're already fully load balancing.
416
417Therefore in the above two situations, the top cpuset flag
418"sched_load_balance" should be disabled, and only some of the smaller,
419child cpusets have this flag enabled.
420
421When doing this, you don't usually want to leave any unpinned tasks in
422the top cpuset that might use non-trivial amounts of CPU, as such tasks
423may be artificially constrained to some subset of CPUs, depending on
424the particulars of this flag setting in descendent cpusets. Even if
425such a task could use spare CPU cycles in some other CPUs, the kernel
426scheduler might not consider the possibility of load balancing that
427task to that underused CPU.
428
429Of course, tasks pinned to a particular CPU can be left in a cpuset
430that disables "sched_load_balance" as those tasks aren't going anywhere
431else anyway.
432
433There is an impedance mismatch here, between cpusets and sched domains.
434Cpusets are hierarchical and nest. Sched domains are flat; they don't
435overlap and each CPU is in at most one sched domain.
436
437It is necessary for sched domains to be flat because load balancing
438across partially overlapping sets of CPUs would risk unstable dynamics
439that would be beyond our understanding. So if each of two partially
440overlapping cpusets enables the flag 'sched_load_balance', then we
441form a single sched domain that is a superset of both. We won't move
442a task to a CPU outside it cpuset, but the scheduler load balancing
443code might waste some compute cycles considering that possibility.
444
445This mismatch is why there is not a simple one-to-one relation
446between which cpusets have the flag "sched_load_balance" enabled,
447and the sched domain configuration. If a cpuset enables the flag, it
448will get balancing across all its CPUs, but if it disables the flag,
449it will only be assured of no load balancing if no other overlapping
450cpuset enables the flag.
451
452If two cpusets have partially overlapping 'cpus' allowed, and only
453one of them has this flag enabled, then the other may find its
454tasks only partially load balanced, just on the overlapping CPUs.
455This is just the general case of the top_cpuset example given a few
456paragraphs above. In the general case, as in the top cpuset case,
457don't leave tasks that might use non-trivial amounts of CPU in
458such partially load balanced cpusets, as they may be artificially
459constrained to some subset of the CPUs allowed to them, for lack of
460load balancing to the other CPUs.
461
4621.7.1 sched_load_balance implementation details.
463------------------------------------------------
464
465The per-cpuset flag 'sched_load_balance' defaults to enabled (contrary
466to most cpuset flags.) When enabled for a cpuset, the kernel will
467ensure that it can load balance across all the CPUs in that cpuset
468(makes sure that all the CPUs in the cpus_allowed of that cpuset are
469in the same sched domain.)
470
471If two overlapping cpusets both have 'sched_load_balance' enabled,
472then they will be (must be) both in the same sched domain.
473
474If, as is the default, the top cpuset has 'sched_load_balance' enabled,
475then by the above that means there is a single sched domain covering
476the whole system, regardless of any other cpuset settings.
477
478The kernel commits to user space that it will avoid load balancing
479where it can. It will pick as fine a granularity partition of sched
480domains as it can while still providing load balancing for any set
481of CPUs allowed to a cpuset having 'sched_load_balance' enabled.
482
483The internal kernel cpuset to scheduler interface passes from the
484cpuset code to the scheduler code a partition of the load balanced
485CPUs in the system. This partition is a set of subsets (represented
486as an array of cpumask_t) of CPUs, pairwise disjoint, that cover all
487the CPUs that must be load balanced.
488
489Whenever the 'sched_load_balance' flag changes, or CPUs come or go
490from a cpuset with this flag enabled, or a cpuset with this flag
491enabled is removed, the cpuset code builds a new such partition and
492passes it to the scheduler sched domain setup code, to have the sched
493domains rebuilt as necessary.
494
495This partition exactly defines what sched domains the scheduler should
496setup - one sched domain for each element (cpumask_t) in the partition.
497
498The scheduler remembers the currently active sched domain partitions.
499When the scheduler routine partition_sched_domains() is invoked from
500the cpuset code to update these sched domains, it compares the new
501partition requested with the current, and updates its sched domains,
502removing the old and adding the new, for each change.
503
Hidetoshi Seto4d5f3552008-04-15 14:03:17 +0900504
5051.8 What is sched_relax_domain_level ?
506--------------------------------------
507
508In sched domain, the scheduler migrates tasks in 2 ways; periodic load
509balance on tick, and at time of some schedule events.
510
511When a task is woken up, scheduler try to move the task on idle CPU.
512For example, if a task A running on CPU X activates another task B
513on the same CPU X, and if CPU Y is X's sibling and performing idle,
514then scheduler migrate task B to CPU Y so that task B can start on
515CPU Y without waiting task A on CPU X.
516
517And if a CPU run out of tasks in its runqueue, the CPU try to pull
518extra tasks from other busy CPUs to help them before it is going to
519be idle.
520
521Of course it takes some searching cost to find movable tasks and/or
522idle CPUs, the scheduler might not search all CPUs in the domain
523everytime. In fact, in some architectures, the searching ranges on
524events are limited in the same socket or node where the CPU locates,
525while the load balance on tick searchs all.
526
527For example, assume CPU Z is relatively far from CPU X. Even if CPU Z
528is idle while CPU X and the siblings are busy, scheduler can't migrate
529woken task B from X to Z since it is out of its searching range.
530As the result, task B on CPU X need to wait task A or wait load balance
531on the next tick. For some applications in special situation, waiting
5321 tick may be too long.
533
534The 'sched_relax_domain_level' file allows you to request changing
535this searching range as you like. This file takes int value which
536indicates size of searching range in levels ideally as follows,
537otherwise initial value -1 that indicates the cpuset has no request.
538
539 -1 : no request. use system default or follow request of others.
540 0 : no search.
541 1 : search siblings (hyperthreads in a core).
542 2 : search cores in a package.
543 3 : search cpus in a node [= system wide on non-NUMA system]
544 ( 4 : search nodes in a chunk of node [on NUMA system] )
545 ( 5~ : search system wide [on NUMA system])
546
547This file is per-cpuset and affect the sched domain where the cpuset
548belongs to. Therefore if the flag 'sched_load_balance' of a cpuset
549is disabled, then 'sched_relax_domain_level' have no effect since
550there is no sched domain belonging the cpuset.
551
552If multiple cpusets are overlapping and hence they form a single sched
553domain, the largest value among those is used. Be careful, if one
554requests 0 and others are -1 then 0 is used.
555
556Note that modifying this file will have both good and bad effects,
557and whether it is acceptable or not will be depend on your situation.
558Don't modify this file if you are not sure.
559
560If your situation is:
561 - The migration costs between each cpu can be assumed considerably
562 small(for you) due to your special application's behavior or
563 special hardware support for CPU cache etc.
564 - The searching cost doesn't have impact(for you) or you can make
565 the searching cost enough small by managing cpuset to compact etc.
566 - The latency is required even it sacrifices cache hit rate etc.
567then increasing 'sched_relax_domain_level' would benefit you.
568
569
5701.9 How do I use cpusets ?
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571--------------------------
572
573In order to minimize the impact of cpusets on critical kernel
574code, such as the scheduler, and due to the fact that the kernel
575does not support one task updating the memory placement of another
576task directly, the impact on a task of changing its cpuset CPU
577or Memory Node placement, or of changing to which cpuset a task
578is attached, is subtle.
579
580If a cpuset has its Memory Nodes modified, then for each task attached
581to that cpuset, the next time that the kernel attempts to allocate
582a page of memory for that task, the kernel will notice the change
583in the tasks cpuset, and update its per-task memory placement to
584remain within the new cpusets memory placement. If the task was using
585mempolicy MPOL_BIND, and the nodes to which it was bound overlap with
586its new cpuset, then the task will continue to use whatever subset
587of MPOL_BIND nodes are still allowed in the new cpuset. If the task
588was using MPOL_BIND and now none of its MPOL_BIND nodes are allowed
589in the new cpuset, then the task will be essentially treated as if it
590was MPOL_BIND bound to the new cpuset (even though its numa placement,
591as queried by get_mempolicy(), doesn't change). If a task is moved
592from one cpuset to another, then the kernel will adjust the tasks
593memory placement, as above, the next time that the kernel attempts
594to allocate a page of memory for that task.
595
Paul Jackson8f5aa262008-02-07 00:14:48 -0800596If a cpuset has its 'cpus' modified, then each task in that cpuset
597will have its allowed CPU placement changed immediately. Similarly,
598if a tasks pid is written to a cpusets 'tasks' file, in either its
599current cpuset or another cpuset, then its allowed CPU placement is
600changed immediately. If such a task had been bound to some subset
601of its cpuset using the sched_setaffinity() call, the task will be
602allowed to run on any CPU allowed in its new cpuset, negating the
603affect of the prior sched_setaffinity() call.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604
605In summary, the memory placement of a task whose cpuset is changed is
606updated by the kernel, on the next allocation of a page for that task,
607but the processor placement is not updated, until that tasks pid is
608rewritten to the 'tasks' file of its cpuset. This is done to avoid
609impacting the scheduler code in the kernel with a check for changes
610in a tasks processor placement.
611
Paul Jackson45b07ef2006-01-08 01:00:56 -0800612Normally, once a page is allocated (given a physical page
613of main memory) then that page stays on whatever node it
614was allocated, so long as it remains allocated, even if the
615cpusets memory placement policy 'mems' subsequently changes.
616If the cpuset flag file 'memory_migrate' is set true, then when
617tasks are attached to that cpuset, any pages that task had
618allocated to it on nodes in its previous cpuset are migrated
Christoph Lameterb4fb3762006-03-14 19:50:20 -0800619to the tasks new cpuset. The relative placement of the page within
620the cpuset is preserved during these migration operations if possible.
621For example if the page was on the second valid node of the prior cpuset
622then the page will be placed on the second valid node of the new cpuset.
623
Paul Jackson45b07ef2006-01-08 01:00:56 -0800624Also if 'memory_migrate' is set true, then if that cpusets
625'mems' file is modified, pages allocated to tasks in that
626cpuset, that were on nodes in the previous setting of 'mems',
Christoph Lameterb4fb3762006-03-14 19:50:20 -0800627will be moved to nodes in the new setting of 'mems.'
628Pages that were not in the tasks prior cpuset, or in the cpusets
629prior 'mems' setting, will not be moved.
Paul Jackson45b07ef2006-01-08 01:00:56 -0800630
Tobias Klauserd533f672005-09-10 00:26:46 -0700631There is an exception to the above. If hotplug functionality is used
Linus Torvalds1da177e2005-04-16 15:20:36 -0700632to remove all the CPUs that are currently assigned to a cpuset,
633then the kernel will automatically update the cpus_allowed of all
Paul Jacksonb39c4fa2005-05-20 13:59:15 -0700634tasks attached to CPUs in that cpuset to allow all CPUs. When memory
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635hotplug functionality for removing Memory Nodes is available, a
636similar exception is expected to apply there as well. In general,
637the kernel prefers to violate cpuset placement, over starving a task
638that has had all its allowed CPUs or Memory Nodes taken offline. User
639code should reconfigure cpusets to only refer to online CPUs and Memory
640Nodes when using hotplug to add or remove such resources.
641
642There is a second exception to the above. GFP_ATOMIC requests are
643kernel internal allocations that must be satisfied, immediately.
644The kernel may drop some request, in rare cases even panic, if a
645GFP_ATOMIC alloc fails. If the request cannot be satisfied within
646the current tasks cpuset, then we relax the cpuset, and look for
647memory anywhere we can find it. It's better to violate the cpuset
648than stress the kernel.
649
650To start a new job that is to be contained within a cpuset, the steps are:
651
652 1) mkdir /dev/cpuset
Paul Menage8793d852007-10-18 23:39:39 -0700653 2) mount -t cgroup -ocpuset cpuset /dev/cpuset
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654 3) Create the new cpuset by doing mkdir's and write's (or echo's) in
655 the /dev/cpuset virtual file system.
656 4) Start a task that will be the "founding father" of the new job.
657 5) Attach that task to the new cpuset by writing its pid to the
658 /dev/cpuset tasks file for that cpuset.
659 6) fork, exec or clone the job tasks from this founding father task.
660
661For example, the following sequence of commands will setup a cpuset
662named "Charlie", containing just CPUs 2 and 3, and Memory Node 1,
663and then start a subshell 'sh' in that cpuset:
664
Paul Menage8793d852007-10-18 23:39:39 -0700665 mount -t cgroup -ocpuset cpuset /dev/cpuset
Linus Torvalds1da177e2005-04-16 15:20:36 -0700666 cd /dev/cpuset
667 mkdir Charlie
668 cd Charlie
669 /bin/echo 2-3 > cpus
670 /bin/echo 1 > mems
671 /bin/echo $$ > tasks
672 sh
673 # The subshell 'sh' is now running in cpuset Charlie
674 # The next line should display '/Charlie'
675 cat /proc/self/cpuset
676
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677In the future, a C library interface to cpusets will likely be
678available. For now, the only way to query or modify cpusets is
679via the cpuset file system, using the various cd, mkdir, echo, cat,
680rmdir commands from the shell, or their equivalent from C.
681
682The sched_setaffinity calls can also be done at the shell prompt using
683SGI's runon or Robert Love's taskset. The mbind and set_mempolicy
684calls can be done at the shell prompt using the numactl command
685(part of Andi Kleen's numa package).
686
6872. Usage Examples and Syntax
688============================
689
6902.1 Basic Usage
691---------------
692
693Creating, modifying, using the cpusets can be done through the cpuset
694virtual filesystem.
695
696To mount it, type:
Paul Menage8793d852007-10-18 23:39:39 -0700697# mount -t cgroup -o cpuset cpuset /dev/cpuset
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698
699Then under /dev/cpuset you can find a tree that corresponds to the
700tree of the cpusets in the system. For instance, /dev/cpuset
701is the cpuset that holds the whole system.
702
703If you want to create a new cpuset under /dev/cpuset:
704# cd /dev/cpuset
705# mkdir my_cpuset
706
707Now you want to do something with this cpuset.
708# cd my_cpuset
709
710In this directory you can find several files:
711# ls
Paul Menage78608362008-04-29 01:00:26 -0700712cpus cpu_exclusive mems mem_exclusive mem_hardwall tasks
Linus Torvalds1da177e2005-04-16 15:20:36 -0700713
714Reading them will give you information about the state of this cpuset:
715the CPUs and Memory Nodes it can use, the processes that are using
716it, its properties. By writing to these files you can manipulate
717the cpuset.
718
719Set some flags:
720# /bin/echo 1 > cpu_exclusive
721
722Add some cpus:
723# /bin/echo 0-7 > cpus
724
Simon Horman2400ff72007-04-01 23:49:40 -0700725Add some mems:
726# /bin/echo 0-7 > mems
727
Linus Torvalds1da177e2005-04-16 15:20:36 -0700728Now attach your shell to this cpuset:
729# /bin/echo $$ > tasks
730
731You can also create cpusets inside your cpuset by using mkdir in this
732directory.
733# mkdir my_sub_cs
734
735To remove a cpuset, just use rmdir:
736# rmdir my_sub_cs
737This will fail if the cpuset is in use (has cpusets inside, or has
738processes attached).
739
Paul Menage8793d852007-10-18 23:39:39 -0700740Note that for legacy reasons, the "cpuset" filesystem exists as a
741wrapper around the cgroup filesystem.
742
743The command
744
745mount -t cpuset X /dev/cpuset
746
747is equivalent to
748
749mount -t cgroup -ocpuset X /dev/cpuset
750echo "/sbin/cpuset_release_agent" > /dev/cpuset/release_agent
751
Linus Torvalds1da177e2005-04-16 15:20:36 -07007522.2 Adding/removing cpus
753------------------------
754
755This is the syntax to use when writing in the cpus or mems files
756in cpuset directories:
757
758# /bin/echo 1-4 > cpus -> set cpus list to cpus 1,2,3,4
759# /bin/echo 1,2,3,4 > cpus -> set cpus list to cpus 1,2,3,4
760
7612.3 Setting flags
762-----------------
763
764The syntax is very simple:
765
766# /bin/echo 1 > cpu_exclusive -> set flag 'cpu_exclusive'
767# /bin/echo 0 > cpu_exclusive -> unset flag 'cpu_exclusive'
768
7692.4 Attaching processes
770-----------------------
771
772# /bin/echo PID > tasks
773
774Note that it is PID, not PIDs. You can only attach ONE task at a time.
775If you have several tasks to attach, you have to do it one after another:
776
777# /bin/echo PID1 > tasks
778# /bin/echo PID2 > tasks
779 ...
780# /bin/echo PIDn > tasks
781
782
7833. Questions
784============
785
786Q: what's up with this '/bin/echo' ?
787A: bash's builtin 'echo' command does not check calls to write() against
788 errors. If you use it in the cpuset file system, you won't be
789 able to tell whether a command succeeded or failed.
790
791Q: When I attach processes, only the first of the line gets really attached !
792A: We can only return one error code per call to write(). So you should also
793 put only ONE pid.
794
7954. Contact
796==========
797
798Web: http://www.bullopensource.org/cpuset