cpusets.rst 37 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879
  1. .. _cpusets:
  2. =======
  3. CPUSETS
  4. =======
  5. Copyright (C) 2004 BULL SA.
  6. Written by Simon.Derr@bull.net
  7. - Portions Copyright (c) 2004-2006 Silicon Graphics, Inc.
  8. - Modified by Paul Jackson <pj@sgi.com>
  9. - Modified by Christoph Lameter <cl@linux.com>
  10. - Modified by Paul Menage <menage@google.com>
  11. - Modified by Hidetoshi Seto <seto.hidetoshi@jp.fujitsu.com>
  12. .. CONTENTS:
  13. 1. Cpusets
  14. 1.1 What are cpusets ?
  15. 1.2 Why are cpusets needed ?
  16. 1.3 How are cpusets implemented ?
  17. 1.4 What are exclusive cpusets ?
  18. 1.5 What is memory_pressure ?
  19. 1.6 What is memory spread ?
  20. 1.7 What is sched_load_balance ?
  21. 1.8 What is sched_relax_domain_level ?
  22. 1.9 How do I use cpusets ?
  23. 2. Usage Examples and Syntax
  24. 2.1 Basic Usage
  25. 2.2 Adding/removing cpus
  26. 2.3 Setting flags
  27. 2.4 Attaching processes
  28. 3. Questions
  29. 4. Contact
  30. 1. Cpusets
  31. ==========
  32. 1.1 What are cpusets ?
  33. ----------------------
  34. Cpusets provide a mechanism for assigning a set of CPUs and Memory
  35. Nodes to a set of tasks. In this document "Memory Node" refers to
  36. an on-line node that contains memory.
  37. Cpusets constrain the CPU and Memory placement of tasks to only
  38. the resources within a task's current cpuset. They form a nested
  39. hierarchy visible in a virtual file system. These are the essential
  40. hooks, beyond what is already present, required to manage dynamic
  41. job placement on large systems.
  42. Cpusets use the generic cgroup subsystem described in
  43. Documentation/admin-guide/cgroup-v1/cgroups.rst.
  44. Requests by a task, using the sched_setaffinity(2) system call to
  45. include CPUs in its CPU affinity mask, and using the mbind(2) and
  46. set_mempolicy(2) system calls to include Memory Nodes in its memory
  47. policy, are both filtered through that task's cpuset, filtering out any
  48. CPUs or Memory Nodes not in that cpuset. The scheduler will not
  49. schedule a task on a CPU that is not allowed in its cpus_allowed
  50. vector, and the kernel page allocator will not allocate a page on a
  51. node that is not allowed in the requesting task's mems_allowed vector.
  52. User level code may create and destroy cpusets by name in the cgroup
  53. virtual file system, manage the attributes and permissions of these
  54. cpusets and which CPUs and Memory Nodes are assigned to each cpuset,
  55. specify and query to which cpuset a task is assigned, and list the
  56. task pids assigned to a cpuset.
  57. 1.2 Why are cpusets needed ?
  58. ----------------------------
  59. The management of large computer systems, with many processors (CPUs),
  60. complex memory cache hierarchies and multiple Memory Nodes having
  61. non-uniform access times (NUMA) presents additional challenges for
  62. the efficient scheduling and memory placement of processes.
  63. Frequently more modest sized systems can be operated with adequate
  64. efficiency just by letting the operating system automatically share
  65. the available CPU and Memory resources amongst the requesting tasks.
  66. But larger systems, which benefit more from careful processor and
  67. memory placement to reduce memory access times and contention,
  68. and which typically represent a larger investment for the customer,
  69. can benefit from explicitly placing jobs on properly sized subsets of
  70. the system.
  71. This can be especially valuable on:
  72. * Web Servers running multiple instances of the same web application,
  73. * Servers running different applications (for instance, a web server
  74. and a database), or
  75. * NUMA systems running large HPC applications with demanding
  76. performance characteristics.
  77. These subsets, or "soft partitions" must be able to be dynamically
  78. adjusted, as the job mix changes, without impacting other concurrently
  79. executing jobs. The location of the running jobs pages may also be moved
  80. when the memory locations are changed.
  81. The kernel cpuset patch provides the minimum essential kernel
  82. mechanisms required to efficiently implement such subsets. It
  83. leverages existing CPU and Memory Placement facilities in the Linux
  84. kernel to avoid any additional impact on the critical scheduler or
  85. memory allocator code.
  86. 1.3 How are cpusets implemented ?
  87. ---------------------------------
  88. Cpusets provide a Linux kernel mechanism to constrain which CPUs and
  89. Memory Nodes are used by a process or set of processes.
  90. The Linux kernel already has a pair of mechanisms to specify on which
  91. CPUs a task may be scheduled (sched_setaffinity) and on which Memory
  92. Nodes it may obtain memory (mbind, set_mempolicy).
  93. Cpusets extends these two mechanisms as follows:
  94. - Cpusets are sets of allowed CPUs and Memory Nodes, known to the
  95. kernel.
  96. - Each task in the system is attached to a cpuset, via a pointer
  97. in the task structure to a reference counted cgroup structure.
  98. - Calls to sched_setaffinity are filtered to just those CPUs
  99. allowed in that task's cpuset.
  100. - Calls to mbind and set_mempolicy are filtered to just
  101. those Memory Nodes allowed in that task's cpuset.
  102. - The root cpuset contains all the systems CPUs and Memory
  103. Nodes.
  104. - For any cpuset, one can define child cpusets containing a subset
  105. of the parents CPU and Memory Node resources.
  106. - The hierarchy of cpusets can be mounted at /dev/cpuset, for
  107. browsing and manipulation from user space.
  108. - A cpuset may be marked exclusive, which ensures that no other
  109. cpuset (except direct ancestors and descendants) may contain
  110. any overlapping CPUs or Memory Nodes.
  111. - You can list all the tasks (by pid) attached to any cpuset.
  112. The implementation of cpusets requires a few, simple hooks
  113. into the rest of the kernel, none in performance critical paths:
  114. - in init/main.c, to initialize the root cpuset at system boot.
  115. - in fork and exit, to attach and detach a task from its cpuset.
  116. - in sched_setaffinity, to mask the requested CPUs by what's
  117. allowed in that task's cpuset.
  118. - in sched.c migrate_live_tasks(), to keep migrating tasks within
  119. the CPUs allowed by their cpuset, if possible.
  120. - in the mbind and set_mempolicy system calls, to mask the requested
  121. Memory Nodes by what's allowed in that task's cpuset.
  122. - in page_alloc.c, to restrict memory to allowed nodes.
  123. - in vmscan.c, to restrict page recovery to the current cpuset.
  124. You should mount the "cgroup" filesystem type in order to enable
  125. browsing and modifying the cpusets presently known to the kernel. No
  126. new system calls are added for cpusets - all support for querying and
  127. modifying cpusets is via this cpuset file system.
  128. The /proc/<pid>/status file for each task has four added lines,
  129. displaying the task's cpus_allowed (on which CPUs it may be scheduled)
  130. and mems_allowed (on which Memory Nodes it may obtain memory),
  131. in the two formats seen in the following example::
  132. Cpus_allowed: ffffffff,ffffffff,ffffffff,ffffffff
  133. Cpus_allowed_list: 0-127
  134. Mems_allowed: ffffffff,ffffffff
  135. Mems_allowed_list: 0-63
  136. Each cpuset is represented by a directory in the cgroup file system
  137. containing (on top of the standard cgroup files) the following
  138. files describing that cpuset:
  139. - cpuset.cpus: list of CPUs in that cpuset
  140. - cpuset.mems: list of Memory Nodes in that cpuset
  141. - cpuset.memory_migrate flag: if set, move pages to cpusets nodes
  142. - cpuset.cpu_exclusive flag: is cpu placement exclusive?
  143. - cpuset.mem_exclusive flag: is memory placement exclusive?
  144. - cpuset.mem_hardwall flag: is memory allocation hardwalled
  145. - cpuset.memory_pressure: measure of how much paging pressure in cpuset
  146. - cpuset.memory_spread_page flag: if set, spread page cache evenly on allowed nodes
  147. - cpuset.memory_spread_slab flag: if set, spread slab cache evenly on allowed nodes
  148. - cpuset.sched_load_balance flag: if set, load balance within CPUs on that cpuset
  149. - cpuset.sched_relax_domain_level: the searching range when migrating tasks
  150. In addition, only the root cpuset has the following file:
  151. - cpuset.memory_pressure_enabled flag: compute memory_pressure?
  152. New cpusets are created using the mkdir system call or shell
  153. command. The properties of a cpuset, such as its flags, allowed
  154. CPUs and Memory Nodes, and attached tasks, are modified by writing
  155. to the appropriate file in that cpusets directory, as listed above.
  156. The named hierarchical structure of nested cpusets allows partitioning
  157. a large system into nested, dynamically changeable, "soft-partitions".
  158. The attachment of each task, automatically inherited at fork by any
  159. children of that task, to a cpuset allows organizing the work load
  160. on a system into related sets of tasks such that each set is constrained
  161. to using the CPUs and Memory Nodes of a particular cpuset. A task
  162. may be re-attached to any other cpuset, if allowed by the permissions
  163. on the necessary cpuset file system directories.
  164. Such management of a system "in the large" integrates smoothly with
  165. the detailed placement done on individual tasks and memory regions
  166. using the sched_setaffinity, mbind and set_mempolicy system calls.
  167. The following rules apply to each cpuset:
  168. - Its CPUs and Memory Nodes must be a subset of its parents.
  169. - It can't be marked exclusive unless its parent is.
  170. - If its cpu or memory is exclusive, they may not overlap any sibling.
  171. These rules, and the natural hierarchy of cpusets, enable efficient
  172. enforcement of the exclusive guarantee, without having to scan all
  173. cpusets every time any of them change to ensure nothing overlaps a
  174. exclusive cpuset. Also, the use of a Linux virtual file system (vfs)
  175. to represent the cpuset hierarchy provides for a familiar permission
  176. and name space for cpusets, with a minimum of additional kernel code.
  177. The cpus and mems files in the root (top_cpuset) cpuset are
  178. read-only. The cpus file automatically tracks the value of
  179. cpu_online_mask using a CPU hotplug notifier, and the mems file
  180. automatically tracks the value of node_states[N_MEMORY]--i.e.,
  181. nodes with memory--using the cpuset_track_online_nodes() hook.
  182. The cpuset.effective_cpus and cpuset.effective_mems files are
  183. normally read-only copies of cpuset.cpus and cpuset.mems files
  184. respectively. If the cpuset cgroup filesystem is mounted with the
  185. special "cpuset_v2_mode" option, the behavior of these files will become
  186. similar to the corresponding files in cpuset v2. In other words, hotplug
  187. events will not change cpuset.cpus and cpuset.mems. Those events will
  188. only affect cpuset.effective_cpus and cpuset.effective_mems which show
  189. the actual cpus and memory nodes that are currently used by this cpuset.
  190. See Documentation/admin-guide/cgroup-v2.rst for more information about
  191. cpuset v2 behavior.
  192. 1.4 What are exclusive cpusets ?
  193. --------------------------------
  194. If a cpuset is cpu or mem exclusive, no other cpuset, other than
  195. a direct ancestor or descendant, may share any of the same CPUs or
  196. Memory Nodes.
  197. A cpuset that is cpuset.mem_exclusive *or* cpuset.mem_hardwall is "hardwalled",
  198. i.e. it restricts kernel allocations for page, buffer and other data
  199. commonly shared by the kernel across multiple users. All cpusets,
  200. whether hardwalled or not, restrict allocations of memory for user
  201. space. This enables configuring a system so that several independent
  202. jobs can share common kernel data, such as file system pages, while
  203. isolating each job's user allocation in its own cpuset. To do this,
  204. construct a large mem_exclusive cpuset to hold all the jobs, and
  205. construct child, non-mem_exclusive cpusets for each individual job.
  206. Only a small amount of typical kernel memory, such as requests from
  207. interrupt handlers, is allowed to be taken outside even a
  208. mem_exclusive cpuset.
  209. 1.5 What is memory_pressure ?
  210. -----------------------------
  211. The memory_pressure of a cpuset provides a simple per-cpuset metric
  212. of the rate that the tasks in a cpuset are attempting to free up in
  213. use memory on the nodes of the cpuset to satisfy additional memory
  214. requests.
  215. This enables batch managers monitoring jobs running in dedicated
  216. cpusets to efficiently detect what level of memory pressure that job
  217. is causing.
  218. This is useful both on tightly managed systems running a wide mix of
  219. submitted jobs, which may choose to terminate or re-prioritize jobs that
  220. are trying to use more memory than allowed on the nodes assigned to them,
  221. and with tightly coupled, long running, massively parallel scientific
  222. computing jobs that will dramatically fail to meet required performance
  223. goals if they start to use more memory than allowed to them.
  224. This mechanism provides a very economical way for the batch manager
  225. to monitor a cpuset for signs of memory pressure. It's up to the
  226. batch manager or other user code to decide what to do about it and
  227. take action.
  228. ==>
  229. Unless this feature is enabled by writing "1" to the special file
  230. /dev/cpuset/memory_pressure_enabled, the hook in the rebalance
  231. code of __alloc_pages() for this metric reduces to simply noticing
  232. that the cpuset_memory_pressure_enabled flag is zero. So only
  233. systems that enable this feature will compute the metric.
  234. Why a per-cpuset, running average:
  235. Because this meter is per-cpuset, rather than per-task or mm,
  236. the system load imposed by a batch scheduler monitoring this
  237. metric is sharply reduced on large systems, because a scan of
  238. the tasklist can be avoided on each set of queries.
  239. Because this meter is a running average, instead of an accumulating
  240. counter, a batch scheduler can detect memory pressure with a
  241. single read, instead of having to read and accumulate results
  242. for a period of time.
  243. Because this meter is per-cpuset rather than per-task or mm,
  244. the batch scheduler can obtain the key information, memory
  245. pressure in a cpuset, with a single read, rather than having to
  246. query and accumulate results over all the (dynamically changing)
  247. set of tasks in the cpuset.
  248. A per-cpuset simple digital filter (requires a spinlock and 3 words
  249. of data per-cpuset) is kept, and updated by any task attached to that
  250. cpuset, if it enters the synchronous (direct) page reclaim code.
  251. A per-cpuset file provides an integer number representing the recent
  252. (half-life of 10 seconds) rate of direct page reclaims caused by
  253. the tasks in the cpuset, in units of reclaims attempted per second,
  254. times 1000.
  255. 1.6 What is memory spread ?
  256. ---------------------------
  257. There are two boolean flag files per cpuset that control where the
  258. kernel allocates pages for the file system buffers and related in
  259. kernel data structures. They are called 'cpuset.memory_spread_page' and
  260. 'cpuset.memory_spread_slab'.
  261. If the per-cpuset boolean flag file 'cpuset.memory_spread_page' is set, then
  262. the kernel will spread the file system buffers (page cache) evenly
  263. over all the nodes that the faulting task is allowed to use, instead
  264. of preferring to put those pages on the node where the task is running.
  265. If the per-cpuset boolean flag file 'cpuset.memory_spread_slab' is set,
  266. then the kernel will spread some file system related slab caches,
  267. such as for inodes and dentries evenly over all the nodes that the
  268. faulting task is allowed to use, instead of preferring to put those
  269. pages on the node where the task is running.
  270. The setting of these flags does not affect anonymous data segment or
  271. stack segment pages of a task.
  272. By default, both kinds of memory spreading are off, and memory
  273. pages are allocated on the node local to where the task is running,
  274. except perhaps as modified by the task's NUMA mempolicy or cpuset
  275. configuration, so long as sufficient free memory pages are available.
  276. When new cpusets are created, they inherit the memory spread settings
  277. of their parent.
  278. Setting memory spreading causes allocations for the affected page
  279. or slab caches to ignore the task's NUMA mempolicy and be spread
  280. instead. Tasks using mbind() or set_mempolicy() calls to set NUMA
  281. mempolicies will not notice any change in these calls as a result of
  282. their containing task's memory spread settings. If memory spreading
  283. is turned off, then the currently specified NUMA mempolicy once again
  284. applies to memory page allocations.
  285. Both 'cpuset.memory_spread_page' and 'cpuset.memory_spread_slab' are boolean flag
  286. files. By default they contain "0", meaning that the feature is off
  287. for that cpuset. If a "1" is written to that file, then that turns
  288. the named feature on.
  289. The implementation is simple.
  290. Setting the flag 'cpuset.memory_spread_page' turns on a per-process flag
  291. PFA_SPREAD_PAGE for each task that is in that cpuset or subsequently
  292. joins that cpuset. The page allocation calls for the page cache
  293. is modified to perform an inline check for this PFA_SPREAD_PAGE task
  294. flag, and if set, a call to a new routine cpuset_mem_spread_node()
  295. returns the node to prefer for the allocation.
  296. Similarly, setting 'cpuset.memory_spread_slab' turns on the flag
  297. PFA_SPREAD_SLAB, and appropriately marked slab caches will allocate
  298. pages from the node returned by cpuset_mem_spread_node().
  299. The cpuset_mem_spread_node() routine is also simple. It uses the
  300. value of a per-task rotor cpuset_mem_spread_rotor to select the next
  301. node in the current task's mems_allowed to prefer for the allocation.
  302. This memory placement policy is also known (in other contexts) as
  303. round-robin or interleave.
  304. This policy can provide substantial improvements for jobs that need
  305. to place thread local data on the corresponding node, but that need
  306. to access large file system data sets that need to be spread across
  307. the several nodes in the jobs cpuset in order to fit. Without this
  308. policy, especially for jobs that might have one thread reading in the
  309. data set, the memory allocation across the nodes in the jobs cpuset
  310. can become very uneven.
  311. 1.7 What is sched_load_balance ?
  312. --------------------------------
  313. The kernel scheduler (kernel/sched/core.c) automatically load balances
  314. tasks. If one CPU is underutilized, kernel code running on that
  315. CPU will look for tasks on other more overloaded CPUs and move those
  316. tasks to itself, within the constraints of such placement mechanisms
  317. as cpusets and sched_setaffinity.
  318. The algorithmic cost of load balancing and its impact on key shared
  319. kernel data structures such as the task list increases more than
  320. linearly with the number of CPUs being balanced. So the scheduler
  321. has support to partition the systems CPUs into a number of sched
  322. domains such that it only load balances within each sched domain.
  323. Each sched domain covers some subset of the CPUs in the system;
  324. no two sched domains overlap; some CPUs might not be in any sched
  325. domain and hence won't be load balanced.
  326. Put simply, it costs less to balance between two smaller sched domains
  327. than one big one, but doing so means that overloads in one of the
  328. two domains won't be load balanced to the other one.
  329. By default, there is one sched domain covering all CPUs, including those
  330. marked isolated using the kernel boot time "isolcpus=" argument. However,
  331. the isolated CPUs will not participate in load balancing, and will not
  332. have tasks running on them unless explicitly assigned.
  333. This default load balancing across all CPUs is not well suited for
  334. the following two situations:
  335. 1) On large systems, load balancing across many CPUs is expensive.
  336. If the system is managed using cpusets to place independent jobs
  337. on separate sets of CPUs, full load balancing is unnecessary.
  338. 2) Systems supporting realtime on some CPUs need to minimize
  339. system overhead on those CPUs, including avoiding task load
  340. balancing if that is not needed.
  341. When the per-cpuset flag "cpuset.sched_load_balance" is enabled (the default
  342. setting), it requests that all the CPUs in that cpusets allowed 'cpuset.cpus'
  343. be contained in a single sched domain, ensuring that load balancing
  344. can move a task (not otherwised pinned, as by sched_setaffinity)
  345. from any CPU in that cpuset to any other.
  346. When the per-cpuset flag "cpuset.sched_load_balance" is disabled, then the
  347. scheduler will avoid load balancing across the CPUs in that cpuset,
  348. --except-- in so far as is necessary because some overlapping cpuset
  349. has "sched_load_balance" enabled.
  350. So, for example, if the top cpuset has the flag "cpuset.sched_load_balance"
  351. enabled, then the scheduler will have one sched domain covering all
  352. CPUs, and the setting of the "cpuset.sched_load_balance" flag in any other
  353. cpusets won't matter, as we're already fully load balancing.
  354. Therefore in the above two situations, the top cpuset flag
  355. "cpuset.sched_load_balance" should be disabled, and only some of the smaller,
  356. child cpusets have this flag enabled.
  357. When doing this, you don't usually want to leave any unpinned tasks in
  358. the top cpuset that might use non-trivial amounts of CPU, as such tasks
  359. may be artificially constrained to some subset of CPUs, depending on
  360. the particulars of this flag setting in descendant cpusets. Even if
  361. such a task could use spare CPU cycles in some other CPUs, the kernel
  362. scheduler might not consider the possibility of load balancing that
  363. task to that underused CPU.
  364. Of course, tasks pinned to a particular CPU can be left in a cpuset
  365. that disables "cpuset.sched_load_balance" as those tasks aren't going anywhere
  366. else anyway.
  367. There is an impedance mismatch here, between cpusets and sched domains.
  368. Cpusets are hierarchical and nest. Sched domains are flat; they don't
  369. overlap and each CPU is in at most one sched domain.
  370. It is necessary for sched domains to be flat because load balancing
  371. across partially overlapping sets of CPUs would risk unstable dynamics
  372. that would be beyond our understanding. So if each of two partially
  373. overlapping cpusets enables the flag 'cpuset.sched_load_balance', then we
  374. form a single sched domain that is a superset of both. We won't move
  375. a task to a CPU outside its cpuset, but the scheduler load balancing
  376. code might waste some compute cycles considering that possibility.
  377. This mismatch is why there is not a simple one-to-one relation
  378. between which cpusets have the flag "cpuset.sched_load_balance" enabled,
  379. and the sched domain configuration. If a cpuset enables the flag, it
  380. will get balancing across all its CPUs, but if it disables the flag,
  381. it will only be assured of no load balancing if no other overlapping
  382. cpuset enables the flag.
  383. If two cpusets have partially overlapping 'cpuset.cpus' allowed, and only
  384. one of them has this flag enabled, then the other may find its
  385. tasks only partially load balanced, just on the overlapping CPUs.
  386. This is just the general case of the top_cpuset example given a few
  387. paragraphs above. In the general case, as in the top cpuset case,
  388. don't leave tasks that might use non-trivial amounts of CPU in
  389. such partially load balanced cpusets, as they may be artificially
  390. constrained to some subset of the CPUs allowed to them, for lack of
  391. load balancing to the other CPUs.
  392. CPUs in "cpuset.isolcpus" were excluded from load balancing by the
  393. isolcpus= kernel boot option, and will never be load balanced regardless
  394. of the value of "cpuset.sched_load_balance" in any cpuset.
  395. 1.7.1 sched_load_balance implementation details.
  396. ------------------------------------------------
  397. The per-cpuset flag 'cpuset.sched_load_balance' defaults to enabled (contrary
  398. to most cpuset flags.) When enabled for a cpuset, the kernel will
  399. ensure that it can load balance across all the CPUs in that cpuset
  400. (makes sure that all the CPUs in the cpus_allowed of that cpuset are
  401. in the same sched domain.)
  402. If two overlapping cpusets both have 'cpuset.sched_load_balance' enabled,
  403. then they will be (must be) both in the same sched domain.
  404. If, as is the default, the top cpuset has 'cpuset.sched_load_balance' enabled,
  405. then by the above that means there is a single sched domain covering
  406. the whole system, regardless of any other cpuset settings.
  407. The kernel commits to user space that it will avoid load balancing
  408. where it can. It will pick as fine a granularity partition of sched
  409. domains as it can while still providing load balancing for any set
  410. of CPUs allowed to a cpuset having 'cpuset.sched_load_balance' enabled.
  411. The internal kernel cpuset to scheduler interface passes from the
  412. cpuset code to the scheduler code a partition of the load balanced
  413. CPUs in the system. This partition is a set of subsets (represented
  414. as an array of struct cpumask) of CPUs, pairwise disjoint, that cover
  415. all the CPUs that must be load balanced.
  416. The cpuset code builds a new such partition and passes it to the
  417. scheduler sched domain setup code, to have the sched domains rebuilt
  418. as necessary, whenever:
  419. - the 'cpuset.sched_load_balance' flag of a cpuset with non-empty CPUs changes,
  420. - or CPUs come or go from a cpuset with this flag enabled,
  421. - or 'cpuset.sched_relax_domain_level' value of a cpuset with non-empty CPUs
  422. and with this flag enabled changes,
  423. - or a cpuset with non-empty CPUs and with this flag enabled is removed,
  424. - or a cpu is offlined/onlined.
  425. This partition exactly defines what sched domains the scheduler should
  426. setup - one sched domain for each element (struct cpumask) in the
  427. partition.
  428. The scheduler remembers the currently active sched domain partitions.
  429. When the scheduler routine partition_sched_domains() is invoked from
  430. the cpuset code to update these sched domains, it compares the new
  431. partition requested with the current, and updates its sched domains,
  432. removing the old and adding the new, for each change.
  433. 1.8 What is sched_relax_domain_level ?
  434. --------------------------------------
  435. In sched domain, the scheduler migrates tasks in 2 ways; periodic load
  436. balance on tick, and at time of some schedule events.
  437. When a task is woken up, scheduler try to move the task on idle CPU.
  438. For example, if a task A running on CPU X activates another task B
  439. on the same CPU X, and if CPU Y is X's sibling and performing idle,
  440. then scheduler migrate task B to CPU Y so that task B can start on
  441. CPU Y without waiting task A on CPU X.
  442. And if a CPU run out of tasks in its runqueue, the CPU try to pull
  443. extra tasks from other busy CPUs to help them before it is going to
  444. be idle.
  445. Of course it takes some searching cost to find movable tasks and/or
  446. idle CPUs, the scheduler might not search all CPUs in the domain
  447. every time. In fact, in some architectures, the searching ranges on
  448. events are limited in the same socket or node where the CPU locates,
  449. while the load balance on tick searches all.
  450. For example, assume CPU Z is relatively far from CPU X. Even if CPU Z
  451. is idle while CPU X and the siblings are busy, scheduler can't migrate
  452. woken task B from X to Z since it is out of its searching range.
  453. As the result, task B on CPU X need to wait task A or wait load balance
  454. on the next tick. For some applications in special situation, waiting
  455. 1 tick may be too long.
  456. The 'cpuset.sched_relax_domain_level' file allows you to request changing
  457. this searching range as you like. This file takes int value which
  458. indicates size of searching range in levels ideally as follows,
  459. otherwise initial value -1 that indicates the cpuset has no request.
  460. ====== ===========================================================
  461. -1 no request. use system default or follow request of others.
  462. 0 no search.
  463. 1 search siblings (hyperthreads in a core).
  464. 2 search cores in a package.
  465. 3 search cpus in a node [= system wide on non-NUMA system]
  466. 4 search nodes in a chunk of node [on NUMA system]
  467. 5 search system wide [on NUMA system]
  468. ====== ===========================================================
  469. The system default is architecture dependent. The system default
  470. can be changed using the relax_domain_level= boot parameter.
  471. This file is per-cpuset and affect the sched domain where the cpuset
  472. belongs to. Therefore if the flag 'cpuset.sched_load_balance' of a cpuset
  473. is disabled, then 'cpuset.sched_relax_domain_level' have no effect since
  474. there is no sched domain belonging the cpuset.
  475. If multiple cpusets are overlapping and hence they form a single sched
  476. domain, the largest value among those is used. Be careful, if one
  477. requests 0 and others are -1 then 0 is used.
  478. Note that modifying this file will have both good and bad effects,
  479. and whether it is acceptable or not depends on your situation.
  480. Don't modify this file if you are not sure.
  481. If your situation is:
  482. - The migration costs between each cpu can be assumed considerably
  483. small(for you) due to your special application's behavior or
  484. special hardware support for CPU cache etc.
  485. - The searching cost doesn't have impact(for you) or you can make
  486. the searching cost enough small by managing cpuset to compact etc.
  487. - The latency is required even it sacrifices cache hit rate etc.
  488. then increasing 'sched_relax_domain_level' would benefit you.
  489. 1.9 How do I use cpusets ?
  490. --------------------------
  491. In order to minimize the impact of cpusets on critical kernel
  492. code, such as the scheduler, and due to the fact that the kernel
  493. does not support one task updating the memory placement of another
  494. task directly, the impact on a task of changing its cpuset CPU
  495. or Memory Node placement, or of changing to which cpuset a task
  496. is attached, is subtle.
  497. If a cpuset has its Memory Nodes modified, then for each task attached
  498. to that cpuset, the next time that the kernel attempts to allocate
  499. a page of memory for that task, the kernel will notice the change
  500. in the task's cpuset, and update its per-task memory placement to
  501. remain within the new cpusets memory placement. If the task was using
  502. mempolicy MPOL_BIND, and the nodes to which it was bound overlap with
  503. its new cpuset, then the task will continue to use whatever subset
  504. of MPOL_BIND nodes are still allowed in the new cpuset. If the task
  505. was using MPOL_BIND and now none of its MPOL_BIND nodes are allowed
  506. in the new cpuset, then the task will be essentially treated as if it
  507. was MPOL_BIND bound to the new cpuset (even though its NUMA placement,
  508. as queried by get_mempolicy(), doesn't change). If a task is moved
  509. from one cpuset to another, then the kernel will adjust the task's
  510. memory placement, as above, the next time that the kernel attempts
  511. to allocate a page of memory for that task.
  512. If a cpuset has its 'cpuset.cpus' modified, then each task in that cpuset
  513. will have its allowed CPU placement changed immediately. Similarly,
  514. if a task's pid is written to another cpuset's 'tasks' file, then its
  515. allowed CPU placement is changed immediately. If such a task had been
  516. bound to some subset of its cpuset using the sched_setaffinity() call,
  517. the task will be allowed to run on any CPU allowed in its new cpuset,
  518. negating the effect of the prior sched_setaffinity() call.
  519. In summary, the memory placement of a task whose cpuset is changed is
  520. updated by the kernel, on the next allocation of a page for that task,
  521. and the processor placement is updated immediately.
  522. Normally, once a page is allocated (given a physical page
  523. of main memory) then that page stays on whatever node it
  524. was allocated, so long as it remains allocated, even if the
  525. cpusets memory placement policy 'cpuset.mems' subsequently changes.
  526. If the cpuset flag file 'cpuset.memory_migrate' is set true, then when
  527. tasks are attached to that cpuset, any pages that task had
  528. allocated to it on nodes in its previous cpuset are migrated
  529. to the task's new cpuset. The relative placement of the page within
  530. the cpuset is preserved during these migration operations if possible.
  531. For example if the page was on the second valid node of the prior cpuset
  532. then the page will be placed on the second valid node of the new cpuset.
  533. Also if 'cpuset.memory_migrate' is set true, then if that cpuset's
  534. 'cpuset.mems' file is modified, pages allocated to tasks in that
  535. cpuset, that were on nodes in the previous setting of 'cpuset.mems',
  536. will be moved to nodes in the new setting of 'mems.'
  537. Pages that were not in the task's prior cpuset, or in the cpuset's
  538. prior 'cpuset.mems' setting, will not be moved.
  539. There is an exception to the above. If hotplug functionality is used
  540. to remove all the CPUs that are currently assigned to a cpuset,
  541. then all the tasks in that cpuset will be moved to the nearest ancestor
  542. with non-empty cpus. But the moving of some (or all) tasks might fail if
  543. cpuset is bound with another cgroup subsystem which has some restrictions
  544. on task attaching. In this failing case, those tasks will stay
  545. in the original cpuset, and the kernel will automatically update
  546. their cpus_allowed to allow all online CPUs. When memory hotplug
  547. functionality for removing Memory Nodes is available, a similar exception
  548. is expected to apply there as well. In general, the kernel prefers to
  549. violate cpuset placement, over starving a task that has had all
  550. its allowed CPUs or Memory Nodes taken offline.
  551. There is a second exception to the above. GFP_ATOMIC requests are
  552. kernel internal allocations that must be satisfied, immediately.
  553. The kernel may drop some request, in rare cases even panic, if a
  554. GFP_ATOMIC alloc fails. If the request cannot be satisfied within
  555. the current task's cpuset, then we relax the cpuset, and look for
  556. memory anywhere we can find it. It's better to violate the cpuset
  557. than stress the kernel.
  558. To start a new job that is to be contained within a cpuset, the steps are:
  559. 1) mkdir /sys/fs/cgroup/cpuset
  560. 2) mount -t cgroup -ocpuset cpuset /sys/fs/cgroup/cpuset
  561. 3) Create the new cpuset by doing mkdir's and write's (or echo's) in
  562. the /sys/fs/cgroup/cpuset virtual file system.
  563. 4) Start a task that will be the "founding father" of the new job.
  564. 5) Attach that task to the new cpuset by writing its pid to the
  565. /sys/fs/cgroup/cpuset tasks file for that cpuset.
  566. 6) fork, exec or clone the job tasks from this founding father task.
  567. For example, the following sequence of commands will setup a cpuset
  568. named "Charlie", containing just CPUs 2 and 3, and Memory Node 1,
  569. and then start a subshell 'sh' in that cpuset::
  570. mount -t cgroup -ocpuset cpuset /sys/fs/cgroup/cpuset
  571. cd /sys/fs/cgroup/cpuset
  572. mkdir Charlie
  573. cd Charlie
  574. /bin/echo 2-3 > cpuset.cpus
  575. /bin/echo 1 > cpuset.mems
  576. /bin/echo $$ > tasks
  577. sh
  578. # The subshell 'sh' is now running in cpuset Charlie
  579. # The next line should display '/Charlie'
  580. cat /proc/self/cpuset
  581. There are ways to query or modify cpusets:
  582. - via the cpuset file system directly, using the various cd, mkdir, echo,
  583. cat, rmdir commands from the shell, or their equivalent from C.
  584. - via the C library libcpuset.
  585. - via the C library libcgroup.
  586. (http://sourceforge.net/projects/libcg/)
  587. - via the python application cset.
  588. (http://code.google.com/p/cpuset/)
  589. The sched_setaffinity calls can also be done at the shell prompt using
  590. SGI's runon or Robert Love's taskset. The mbind and set_mempolicy
  591. calls can be done at the shell prompt using the numactl command
  592. (part of Andi Kleen's numa package).
  593. 2. Usage Examples and Syntax
  594. ============================
  595. 2.1 Basic Usage
  596. ---------------
  597. Creating, modifying, using the cpusets can be done through the cpuset
  598. virtual filesystem.
  599. To mount it, type:
  600. # mount -t cgroup -o cpuset cpuset /sys/fs/cgroup/cpuset
  601. Then under /sys/fs/cgroup/cpuset you can find a tree that corresponds to the
  602. tree of the cpusets in the system. For instance, /sys/fs/cgroup/cpuset
  603. is the cpuset that holds the whole system.
  604. If you want to create a new cpuset under /sys/fs/cgroup/cpuset::
  605. # cd /sys/fs/cgroup/cpuset
  606. # mkdir my_cpuset
  607. Now you want to do something with this cpuset::
  608. # cd my_cpuset
  609. In this directory you can find several files::
  610. # ls
  611. cgroup.clone_children cpuset.memory_pressure
  612. cgroup.event_control cpuset.memory_spread_page
  613. cgroup.procs cpuset.memory_spread_slab
  614. cpuset.cpu_exclusive cpuset.mems
  615. cpuset.cpus cpuset.sched_load_balance
  616. cpuset.mem_exclusive cpuset.sched_relax_domain_level
  617. cpuset.mem_hardwall notify_on_release
  618. cpuset.memory_migrate tasks
  619. Reading them will give you information about the state of this cpuset:
  620. the CPUs and Memory Nodes it can use, the processes that are using
  621. it, its properties. By writing to these files you can manipulate
  622. the cpuset.
  623. Set some flags::
  624. # /bin/echo 1 > cpuset.cpu_exclusive
  625. Add some cpus::
  626. # /bin/echo 0-7 > cpuset.cpus
  627. Add some mems::
  628. # /bin/echo 0-7 > cpuset.mems
  629. Now attach your shell to this cpuset::
  630. # /bin/echo $$ > tasks
  631. You can also create cpusets inside your cpuset by using mkdir in this
  632. directory::
  633. # mkdir my_sub_cs
  634. To remove a cpuset, just use rmdir::
  635. # rmdir my_sub_cs
  636. This will fail if the cpuset is in use (has cpusets inside, or has
  637. processes attached).
  638. Note that for legacy reasons, the "cpuset" filesystem exists as a
  639. wrapper around the cgroup filesystem.
  640. The command::
  641. mount -t cpuset X /sys/fs/cgroup/cpuset
  642. is equivalent to::
  643. mount -t cgroup -ocpuset,noprefix X /sys/fs/cgroup/cpuset
  644. echo "/sbin/cpuset_release_agent" > /sys/fs/cgroup/cpuset/release_agent
  645. 2.2 Adding/removing cpus
  646. ------------------------
  647. This is the syntax to use when writing in the cpus or mems files
  648. in cpuset directories::
  649. # /bin/echo 1-4 > cpuset.cpus -> set cpus list to cpus 1,2,3,4
  650. # /bin/echo 1,2,3,4 > cpuset.cpus -> set cpus list to cpus 1,2,3,4
  651. To add a CPU to a cpuset, write the new list of CPUs including the
  652. CPU to be added. To add 6 to the above cpuset::
  653. # /bin/echo 1-4,6 > cpuset.cpus -> set cpus list to cpus 1,2,3,4,6
  654. Similarly to remove a CPU from a cpuset, write the new list of CPUs
  655. without the CPU to be removed.
  656. To remove all the CPUs::
  657. # /bin/echo "" > cpuset.cpus -> clear cpus list
  658. 2.3 Setting flags
  659. -----------------
  660. The syntax is very simple::
  661. # /bin/echo 1 > cpuset.cpu_exclusive -> set flag 'cpuset.cpu_exclusive'
  662. # /bin/echo 0 > cpuset.cpu_exclusive -> unset flag 'cpuset.cpu_exclusive'
  663. 2.4 Attaching processes
  664. -----------------------
  665. ::
  666. # /bin/echo PID > tasks
  667. Note that it is PID, not PIDs. You can only attach ONE task at a time.
  668. If you have several tasks to attach, you have to do it one after another::
  669. # /bin/echo PID1 > tasks
  670. # /bin/echo PID2 > tasks
  671. ...
  672. # /bin/echo PIDn > tasks
  673. 3. Questions
  674. ============
  675. Q:
  676. what's up with this '/bin/echo' ?
  677. A:
  678. bash's builtin 'echo' command does not check calls to write() against
  679. errors. If you use it in the cpuset file system, you won't be
  680. able to tell whether a command succeeded or failed.
  681. Q:
  682. When I attach processes, only the first of the line gets really attached !
  683. A:
  684. We can only return one error code per call to write(). So you should also
  685. put only ONE pid.
  686. 4. Contact
  687. ==========
  688. Web: http://www.bullopensource.org/cpuset