blob: b875d231ac748c64129865b8b4c130c929b474c1 [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>
Linus Torvalds1da177e2005-04-16 15:20:36 -070010
11CONTENTS:
12=========
13
141. Cpusets
15 1.1 What are cpusets ?
16 1.2 Why are cpusets needed ?
17 1.3 How are cpusets implemented ?
Paul Jacksonbd5e09c2006-01-08 01:01:50 -080018 1.4 What are exclusive cpusets ?
19 1.5 What does notify_on_release do ?
Paul Jackson90c9cc42006-01-08 01:01:51 -080020 1.6 What is memory_pressure ?
Paul Jackson825a46a2006-03-24 03:16:03 -080021 1.7 What is memory spread ?
22 1.8 How do I use cpusets ?
Linus Torvalds1da177e2005-04-16 15:20:36 -0700232. 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
283. Questions
294. Contact
30
311. Cpusets
32==========
33
341.1 What are cpusets ?
35----------------------
36
37Cpusets provide a mechanism for assigning a set of CPUs and Memory
Christoph Lameter0e1e7c72007-10-16 01:25:38 -070038Nodes to a set of tasks. In this document "Memory Node" refers to
39an on-line node that contains memory.
Linus Torvalds1da177e2005-04-16 15:20:36 -070040
41Cpusets constrain the CPU and Memory placement of tasks to only
42the resources within a tasks current cpuset. They form a nested
43hierarchy visible in a virtual file system. These are the essential
44hooks, beyond what is already present, required to manage dynamic
45job placement on large systems.
46
47Each task has a pointer to a cpuset. Multiple tasks may reference
48the same cpuset. Requests by a task, using the sched_setaffinity(2)
49system call to include CPUs in its CPU affinity mask, and using the
50mbind(2) and set_mempolicy(2) system calls to include Memory Nodes
51in its memory policy, are both filtered through that tasks cpuset,
52filtering out any CPUs or Memory Nodes not in that cpuset. The
53scheduler will not schedule a task on a CPU that is not allowed in
54its cpus_allowed vector, and the kernel page allocator will not
55allocate a page on a node that is not allowed in the requesting tasks
56mems_allowed vector.
57
Linus Torvalds1da177e2005-04-16 15:20:36 -070058User level code may create and destroy cpusets by name in the cpuset
59virtual file system, manage the attributes and permissions of these
60cpusets and which CPUs and Memory Nodes are assigned to each cpuset,
61specify and query to which cpuset a task is assigned, and list the
62task pids assigned to a cpuset.
63
64
651.2 Why are cpusets needed ?
66----------------------------
67
68The management of large computer systems, with many processors (CPUs),
69complex memory cache hierarchies and multiple Memory Nodes having
70non-uniform access times (NUMA) presents additional challenges for
71the efficient scheduling and memory placement of processes.
72
73Frequently more modest sized systems can be operated with adequate
74efficiency just by letting the operating system automatically share
75the available CPU and Memory resources amongst the requesting tasks.
76
77But larger systems, which benefit more from careful processor and
78memory placement to reduce memory access times and contention,
79and which typically represent a larger investment for the customer,
Jean Delvare33430dc2005-10-30 15:02:20 -080080can benefit from explicitly placing jobs on properly sized subsets of
Linus Torvalds1da177e2005-04-16 15:20:36 -070081the system.
82
83This can be especially valuable on:
84
85 * Web Servers running multiple instances of the same web application,
86 * Servers running different applications (for instance, a web server
87 and a database), or
88 * NUMA systems running large HPC applications with demanding
89 performance characteristics.
Dinakar Guniguntala85d7b942005-06-25 14:57:34 -070090 * Also cpu_exclusive cpusets are useful for servers running orthogonal
91 workloads such as RT applications requiring low latency and HPC
92 applications that are throughput sensitive
Linus Torvalds1da177e2005-04-16 15:20:36 -070093
94These subsets, or "soft partitions" must be able to be dynamically
95adjusted, as the job mix changes, without impacting other concurrently
Christoph Lameterb4fb3762006-03-14 19:50:20 -080096executing jobs. The location of the running jobs pages may also be moved
97when the memory locations are changed.
Linus Torvalds1da177e2005-04-16 15:20:36 -070098
99The kernel cpuset patch provides the minimum essential kernel
100mechanisms required to efficiently implement such subsets. It
101leverages existing CPU and Memory Placement facilities in the Linux
102kernel to avoid any additional impact on the critical scheduler or
103memory allocator code.
104
105
1061.3 How are cpusets implemented ?
107---------------------------------
108
Christoph Lameterb4fb3762006-03-14 19:50:20 -0800109Cpusets provide a Linux kernel mechanism to constrain which CPUs and
110Memory Nodes are used by a process or set of processes.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111
112The Linux kernel already has a pair of mechanisms to specify on which
113CPUs a task may be scheduled (sched_setaffinity) and on which Memory
114Nodes it may obtain memory (mbind, set_mempolicy).
115
116Cpusets extends these two mechanisms as follows:
117
118 - Cpusets are sets of allowed CPUs and Memory Nodes, known to the
119 kernel.
120 - Each task in the system is attached to a cpuset, via a pointer
121 in the task structure to a reference counted cpuset structure.
122 - Calls to sched_setaffinity are filtered to just those CPUs
123 allowed in that tasks cpuset.
124 - Calls to mbind and set_mempolicy are filtered to just
125 those Memory Nodes allowed in that tasks cpuset.
126 - The root cpuset contains all the systems CPUs and Memory
127 Nodes.
128 - For any cpuset, one can define child cpusets containing a subset
129 of the parents CPU and Memory Node resources.
130 - The hierarchy of cpusets can be mounted at /dev/cpuset, for
131 browsing and manipulation from user space.
132 - A cpuset may be marked exclusive, which ensures that no other
133 cpuset (except direct ancestors and descendents) may contain
134 any overlapping CPUs or Memory Nodes.
Dinakar Guniguntala85d7b942005-06-25 14:57:34 -0700135 Also a cpu_exclusive cpuset would be associated with a sched
136 domain.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700137 - You can list all the tasks (by pid) attached to any cpuset.
138
139The implementation of cpusets requires a few, simple hooks
140into the rest of the kernel, none in performance critical paths:
141
Paul Jackson864913f2006-01-11 02:01:38 +0100142 - in init/main.c, to initialize the root cpuset at system boot.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143 - in fork and exit, to attach and detach a task from its cpuset.
144 - in sched_setaffinity, to mask the requested CPUs by what's
145 allowed in that tasks cpuset.
146 - in sched.c migrate_all_tasks(), to keep migrating tasks within
147 the CPUs allowed by their cpuset, if possible.
Dinakar Guniguntala85d7b942005-06-25 14:57:34 -0700148 - in sched.c, a new API partition_sched_domains for handling
149 sched domain changes associated with cpu_exclusive cpusets
150 and related changes in both sched.c and arch/ia64/kernel/domain.c
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151 - in the mbind and set_mempolicy system calls, to mask the requested
152 Memory Nodes by what's allowed in that tasks cpuset.
Paul Jackson864913f2006-01-11 02:01:38 +0100153 - in page_alloc.c, to restrict memory to allowed nodes.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154 - in vmscan.c, to restrict page recovery to the current cpuset.
155
156In addition a new file system, of type "cpuset" may be mounted,
157typically at /dev/cpuset, to enable browsing and modifying the cpusets
158presently known to the kernel. No new system calls are added for
159cpusets - all support for querying and modifying cpusets is via
160this cpuset file system.
161
162Each task under /proc has an added file named 'cpuset', displaying
163the cpuset name, as the path relative to the root of the cpuset file
164system.
165
166The /proc/<pid>/status file for each task has two added lines,
167displaying the tasks cpus_allowed (on which CPUs it may be scheduled)
168and mems_allowed (on which Memory Nodes it may obtain memory),
169in the format seen in the following example:
170
171 Cpus_allowed: ffffffff,ffffffff,ffffffff,ffffffff
172 Mems_allowed: ffffffff,ffffffff
173
174Each cpuset is represented by a directory in the cpuset file system
175containing the following files describing that cpuset:
176
177 - cpus: list of CPUs in that cpuset
178 - mems: list of Memory Nodes in that cpuset
Paul Jackson45b07ef2006-01-08 01:00:56 -0800179 - memory_migrate flag: if set, move pages to cpusets nodes
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180 - cpu_exclusive flag: is cpu placement exclusive?
181 - mem_exclusive flag: is memory placement exclusive?
182 - tasks: list of tasks (by pid) attached to that cpuset
Paul Jacksonbd5e09c2006-01-08 01:01:50 -0800183 - notify_on_release flag: run /sbin/cpuset_release_agent on exit?
Paul Jacksonbd5e09c2006-01-08 01:01:50 -0800184 - memory_pressure: measure of how much paging pressure in cpuset
185
186In addition, the root cpuset only has the following file:
187 - memory_pressure_enabled flag: compute memory_pressure?
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188
189New cpusets are created using the mkdir system call or shell
190command. The properties of a cpuset, such as its flags, allowed
191CPUs and Memory Nodes, and attached tasks, are modified by writing
192to the appropriate file in that cpusets directory, as listed above.
193
194The named hierarchical structure of nested cpusets allows partitioning
195a large system into nested, dynamically changeable, "soft-partitions".
196
197The attachment of each task, automatically inherited at fork by any
198children of that task, to a cpuset allows organizing the work load
199on a system into related sets of tasks such that each set is constrained
200to using the CPUs and Memory Nodes of a particular cpuset. A task
201may be re-attached to any other cpuset, if allowed by the permissions
202on the necessary cpuset file system directories.
203
204Such management of a system "in the large" integrates smoothly with
205the detailed placement done on individual tasks and memory regions
206using the sched_setaffinity, mbind and set_mempolicy system calls.
207
208The following rules apply to each cpuset:
209
210 - Its CPUs and Memory Nodes must be a subset of its parents.
211 - It can only be marked exclusive if its parent is.
212 - If its cpu or memory is exclusive, they may not overlap any sibling.
213
214These rules, and the natural hierarchy of cpusets, enable efficient
215enforcement of the exclusive guarantee, without having to scan all
216cpusets every time any of them change to ensure nothing overlaps a
217exclusive cpuset. Also, the use of a Linux virtual file system (vfs)
218to represent the cpuset hierarchy provides for a familiar permission
219and name space for cpusets, with a minimum of additional kernel code.
220
Paul Jackson38837fc2006-09-29 02:01:16 -0700221The cpus and mems files in the root (top_cpuset) cpuset are
222read-only. The cpus file automatically tracks the value of
223cpu_online_map using a CPU hotplug notifier, and the mems file
Christoph Lameter0e1e7c72007-10-16 01:25:38 -0700224automatically tracks the value of node_states[N_MEMORY]--i.e.,
225nodes with memory--using the cpuset_track_online_nodes() hook.
Paul Jackson4c4d50f2006-08-27 01:23:51 -0700226
Paul Jacksonbd5e09c2006-01-08 01:01:50 -0800227
2281.4 What are exclusive cpusets ?
229--------------------------------
230
231If a cpuset is cpu or mem exclusive, no other cpuset, other than
232a direct ancestor or descendent, may share any of the same CPUs or
233Memory Nodes.
234
235A cpuset that is cpu_exclusive has a scheduler (sched) domain
236associated with it. The sched domain consists of all CPUs in the
237current cpuset that are not part of any exclusive child cpusets.
238This ensures that the scheduler load balancing code only balances
239against the CPUs that are in the sched domain as defined above and
240not all of the CPUs in the system. This removes any overhead due to
241load balancing code trying to pull tasks outside of the cpu_exclusive
242cpuset only to be prevented by the tasks' cpus_allowed mask.
243
244A cpuset that is mem_exclusive restricts kernel allocations for
245page, buffer and other data commonly shared by the kernel across
246multiple users. All cpusets, whether mem_exclusive or not, restrict
247allocations of memory for user space. This enables configuring a
248system so that several independent jobs can share common kernel data,
249such as file system pages, while isolating each jobs user allocation in
250its own cpuset. To do this, construct a large mem_exclusive cpuset to
251hold all the jobs, and construct child, non-mem_exclusive cpusets for
252each individual job. Only a small amount of typical kernel memory,
253such as requests from interrupt handlers, is allowed to be taken
254outside even a mem_exclusive cpuset.
255
256
2571.5 What does notify_on_release do ?
258------------------------------------
259
260If the notify_on_release flag is enabled (1) in a cpuset, then whenever
261the last task in the cpuset leaves (exits or attaches to some other
262cpuset) and the last child cpuset of that cpuset is removed, then
263the kernel runs the command /sbin/cpuset_release_agent, supplying the
264pathname (relative to the mount point of the cpuset file system) of the
265abandoned cpuset. This enables automatic removal of abandoned cpusets.
266The default value of notify_on_release in the root cpuset at system
267boot is disabled (0). The default value of other cpusets at creation
268is the current value of their parents notify_on_release setting.
269
270
Paul Jackson90c9cc42006-01-08 01:01:51 -08002711.6 What is memory_pressure ?
Paul Jacksonbd5e09c2006-01-08 01:01:50 -0800272-----------------------------
273The memory_pressure of a cpuset provides a simple per-cpuset metric
274of the rate that the tasks in a cpuset are attempting to free up in
275use memory on the nodes of the cpuset to satisfy additional memory
276requests.
277
278This enables batch managers monitoring jobs running in dedicated
279cpusets to efficiently detect what level of memory pressure that job
280is causing.
281
282This is useful both on tightly managed systems running a wide mix of
283submitted jobs, which may choose to terminate or re-prioritize jobs that
284are trying to use more memory than allowed on the nodes assigned them,
285and with tightly coupled, long running, massively parallel scientific
286computing jobs that will dramatically fail to meet required performance
287goals if they start to use more memory than allowed to them.
288
289This mechanism provides a very economical way for the batch manager
290to monitor a cpuset for signs of memory pressure. It's up to the
291batch manager or other user code to decide what to do about it and
292take action.
293
294==> Unless this feature is enabled by writing "1" to the special file
295 /dev/cpuset/memory_pressure_enabled, the hook in the rebalance
296 code of __alloc_pages() for this metric reduces to simply noticing
297 that the cpuset_memory_pressure_enabled flag is zero. So only
298 systems that enable this feature will compute the metric.
299
300Why a per-cpuset, running average:
301
302 Because this meter is per-cpuset, rather than per-task or mm,
303 the system load imposed by a batch scheduler monitoring this
304 metric is sharply reduced on large systems, because a scan of
305 the tasklist can be avoided on each set of queries.
306
307 Because this meter is a running average, instead of an accumulating
308 counter, a batch scheduler can detect memory pressure with a
309 single read, instead of having to read and accumulate results
310 for a period of time.
311
312 Because this meter is per-cpuset rather than per-task or mm,
313 the batch scheduler can obtain the key information, memory
314 pressure in a cpuset, with a single read, rather than having to
315 query and accumulate results over all the (dynamically changing)
316 set of tasks in the cpuset.
317
318A per-cpuset simple digital filter (requires a spinlock and 3 words
319of data per-cpuset) is kept, and updated by any task attached to that
320cpuset, if it enters the synchronous (direct) page reclaim code.
321
322A per-cpuset file provides an integer number representing the recent
323(half-life of 10 seconds) rate of direct page reclaims caused by
324the tasks in the cpuset, in units of reclaims attempted per second,
325times 1000.
326
327
Paul Jackson825a46a2006-03-24 03:16:03 -08003281.7 What is memory spread ?
329---------------------------
330There are two boolean flag files per cpuset that control where the
331kernel allocates pages for the file system buffers and related in
332kernel data structures. They are called 'memory_spread_page' and
333'memory_spread_slab'.
334
335If the per-cpuset boolean flag file 'memory_spread_page' is set, then
336the kernel will spread the file system buffers (page cache) evenly
337over all the nodes that the faulting task is allowed to use, instead
338of preferring to put those pages on the node where the task is running.
339
340If the per-cpuset boolean flag file 'memory_spread_slab' is set,
341then the kernel will spread some file system related slab caches,
342such as for inodes and dentries evenly over all the nodes that the
343faulting task is allowed to use, instead of preferring to put those
344pages on the node where the task is running.
345
346The setting of these flags does not affect anonymous data segment or
347stack segment pages of a task.
348
349By default, both kinds of memory spreading are off, and memory
350pages are allocated on the node local to where the task is running,
351except perhaps as modified by the tasks NUMA mempolicy or cpuset
352configuration, so long as sufficient free memory pages are available.
353
354When new cpusets are created, they inherit the memory spread settings
355of their parent.
356
357Setting memory spreading causes allocations for the affected page
358or slab caches to ignore the tasks NUMA mempolicy and be spread
359instead. Tasks using mbind() or set_mempolicy() calls to set NUMA
360mempolicies will not notice any change in these calls as a result of
361their containing tasks memory spread settings. If memory spreading
362is turned off, then the currently specified NUMA mempolicy once again
363applies to memory page allocations.
364
365Both 'memory_spread_page' and 'memory_spread_slab' are boolean flag
366files. By default they contain "0", meaning that the feature is off
367for that cpuset. If a "1" is written to that file, then that turns
368the named feature on.
369
370The implementation is simple.
371
372Setting the flag 'memory_spread_page' turns on a per-process flag
373PF_SPREAD_PAGE for each task that is in that cpuset or subsequently
374joins that cpuset. The page allocation calls for the page cache
375is modified to perform an inline check for this PF_SPREAD_PAGE task
376flag, and if set, a call to a new routine cpuset_mem_spread_node()
377returns the node to prefer for the allocation.
378
379Similarly, setting 'memory_spread_cache' turns on the flag
380PF_SPREAD_SLAB, and appropriately marked slab caches will allocate
381pages from the node returned by cpuset_mem_spread_node().
382
383The cpuset_mem_spread_node() routine is also simple. It uses the
384value of a per-task rotor cpuset_mem_spread_rotor to select the next
385node in the current tasks mems_allowed to prefer for the allocation.
386
387This memory placement policy is also known (in other contexts) as
388round-robin or interleave.
389
390This policy can provide substantial improvements for jobs that need
391to place thread local data on the corresponding node, but that need
392to access large file system data sets that need to be spread across
393the several nodes in the jobs cpuset in order to fit. Without this
394policy, especially for jobs that might have one thread reading in the
395data set, the memory allocation across the nodes in the jobs cpuset
396can become very uneven.
397
398
3991.8 How do I use cpusets ?
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400--------------------------
401
402In order to minimize the impact of cpusets on critical kernel
403code, such as the scheduler, and due to the fact that the kernel
404does not support one task updating the memory placement of another
405task directly, the impact on a task of changing its cpuset CPU
406or Memory Node placement, or of changing to which cpuset a task
407is attached, is subtle.
408
409If a cpuset has its Memory Nodes modified, then for each task attached
410to that cpuset, the next time that the kernel attempts to allocate
411a page of memory for that task, the kernel will notice the change
412in the tasks cpuset, and update its per-task memory placement to
413remain within the new cpusets memory placement. If the task was using
414mempolicy MPOL_BIND, and the nodes to which it was bound overlap with
415its new cpuset, then the task will continue to use whatever subset
416of MPOL_BIND nodes are still allowed in the new cpuset. If the task
417was using MPOL_BIND and now none of its MPOL_BIND nodes are allowed
418in the new cpuset, then the task will be essentially treated as if it
419was MPOL_BIND bound to the new cpuset (even though its numa placement,
420as queried by get_mempolicy(), doesn't change). If a task is moved
421from one cpuset to another, then the kernel will adjust the tasks
422memory placement, as above, the next time that the kernel attempts
423to allocate a page of memory for that task.
424
425If a cpuset has its CPUs modified, then each task using that
426cpuset does _not_ change its behavior automatically. In order to
427minimize the impact on the critical scheduling code in the kernel,
428tasks will continue to use their prior CPU placement until they
429are rebound to their cpuset, by rewriting their pid to the 'tasks'
430file of their cpuset. If a task had been bound to some subset of its
431cpuset using the sched_setaffinity() call, and if any of that subset
432is still allowed in its new cpuset settings, then the task will be
433restricted to the intersection of the CPUs it was allowed on before,
434and its new cpuset CPU placement. If, on the other hand, there is
435no overlap between a tasks prior placement and its new cpuset CPU
436placement, then the task will be allowed to run on any CPU allowed
437in its new cpuset. If a task is moved from one cpuset to another,
438its CPU placement is updated in the same way as if the tasks pid is
439rewritten to the 'tasks' file of its current cpuset.
440
441In summary, the memory placement of a task whose cpuset is changed is
442updated by the kernel, on the next allocation of a page for that task,
443but the processor placement is not updated, until that tasks pid is
444rewritten to the 'tasks' file of its cpuset. This is done to avoid
445impacting the scheduler code in the kernel with a check for changes
446in a tasks processor placement.
447
Paul Jackson45b07ef2006-01-08 01:00:56 -0800448Normally, once a page is allocated (given a physical page
449of main memory) then that page stays on whatever node it
450was allocated, so long as it remains allocated, even if the
451cpusets memory placement policy 'mems' subsequently changes.
452If the cpuset flag file 'memory_migrate' is set true, then when
453tasks are attached to that cpuset, any pages that task had
454allocated to it on nodes in its previous cpuset are migrated
Christoph Lameterb4fb3762006-03-14 19:50:20 -0800455to the tasks new cpuset. The relative placement of the page within
456the cpuset is preserved during these migration operations if possible.
457For example if the page was on the second valid node of the prior cpuset
458then the page will be placed on the second valid node of the new cpuset.
459
Paul Jackson45b07ef2006-01-08 01:00:56 -0800460Also if 'memory_migrate' is set true, then if that cpusets
461'mems' file is modified, pages allocated to tasks in that
462cpuset, that were on nodes in the previous setting of 'mems',
Christoph Lameterb4fb3762006-03-14 19:50:20 -0800463will be moved to nodes in the new setting of 'mems.'
464Pages that were not in the tasks prior cpuset, or in the cpusets
465prior 'mems' setting, will not be moved.
Paul Jackson45b07ef2006-01-08 01:00:56 -0800466
Tobias Klauserd533f672005-09-10 00:26:46 -0700467There is an exception to the above. If hotplug functionality is used
Linus Torvalds1da177e2005-04-16 15:20:36 -0700468to remove all the CPUs that are currently assigned to a cpuset,
469then the kernel will automatically update the cpus_allowed of all
Paul Jacksonb39c4fa2005-05-20 13:59:15 -0700470tasks attached to CPUs in that cpuset to allow all CPUs. When memory
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471hotplug functionality for removing Memory Nodes is available, a
472similar exception is expected to apply there as well. In general,
473the kernel prefers to violate cpuset placement, over starving a task
474that has had all its allowed CPUs or Memory Nodes taken offline. User
475code should reconfigure cpusets to only refer to online CPUs and Memory
476Nodes when using hotplug to add or remove such resources.
477
478There is a second exception to the above. GFP_ATOMIC requests are
479kernel internal allocations that must be satisfied, immediately.
480The kernel may drop some request, in rare cases even panic, if a
481GFP_ATOMIC alloc fails. If the request cannot be satisfied within
482the current tasks cpuset, then we relax the cpuset, and look for
483memory anywhere we can find it. It's better to violate the cpuset
484than stress the kernel.
485
486To start a new job that is to be contained within a cpuset, the steps are:
487
488 1) mkdir /dev/cpuset
489 2) mount -t cpuset none /dev/cpuset
490 3) Create the new cpuset by doing mkdir's and write's (or echo's) in
491 the /dev/cpuset virtual file system.
492 4) Start a task that will be the "founding father" of the new job.
493 5) Attach that task to the new cpuset by writing its pid to the
494 /dev/cpuset tasks file for that cpuset.
495 6) fork, exec or clone the job tasks from this founding father task.
496
497For example, the following sequence of commands will setup a cpuset
498named "Charlie", containing just CPUs 2 and 3, and Memory Node 1,
499and then start a subshell 'sh' in that cpuset:
500
501 mount -t cpuset none /dev/cpuset
502 cd /dev/cpuset
503 mkdir Charlie
504 cd Charlie
505 /bin/echo 2-3 > cpus
506 /bin/echo 1 > mems
507 /bin/echo $$ > tasks
508 sh
509 # The subshell 'sh' is now running in cpuset Charlie
510 # The next line should display '/Charlie'
511 cat /proc/self/cpuset
512
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513In the future, a C library interface to cpusets will likely be
514available. For now, the only way to query or modify cpusets is
515via the cpuset file system, using the various cd, mkdir, echo, cat,
516rmdir commands from the shell, or their equivalent from C.
517
518The sched_setaffinity calls can also be done at the shell prompt using
519SGI's runon or Robert Love's taskset. The mbind and set_mempolicy
520calls can be done at the shell prompt using the numactl command
521(part of Andi Kleen's numa package).
522
5232. Usage Examples and Syntax
524============================
525
5262.1 Basic Usage
527---------------
528
529Creating, modifying, using the cpusets can be done through the cpuset
530virtual filesystem.
531
532To mount it, type:
533# mount -t cpuset none /dev/cpuset
534
535Then under /dev/cpuset you can find a tree that corresponds to the
536tree of the cpusets in the system. For instance, /dev/cpuset
537is the cpuset that holds the whole system.
538
539If you want to create a new cpuset under /dev/cpuset:
540# cd /dev/cpuset
541# mkdir my_cpuset
542
543Now you want to do something with this cpuset.
544# cd my_cpuset
545
546In this directory you can find several files:
547# ls
548cpus cpu_exclusive mems mem_exclusive tasks
549
550Reading them will give you information about the state of this cpuset:
551the CPUs and Memory Nodes it can use, the processes that are using
552it, its properties. By writing to these files you can manipulate
553the cpuset.
554
555Set some flags:
556# /bin/echo 1 > cpu_exclusive
557
558Add some cpus:
559# /bin/echo 0-7 > cpus
560
Simon Horman2400ff72007-04-01 23:49:40 -0700561Add some mems:
562# /bin/echo 0-7 > mems
563
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564Now attach your shell to this cpuset:
565# /bin/echo $$ > tasks
566
567You can also create cpusets inside your cpuset by using mkdir in this
568directory.
569# mkdir my_sub_cs
570
571To remove a cpuset, just use rmdir:
572# rmdir my_sub_cs
573This will fail if the cpuset is in use (has cpusets inside, or has
574processes attached).
575
5762.2 Adding/removing cpus
577------------------------
578
579This is the syntax to use when writing in the cpus or mems files
580in cpuset directories:
581
582# /bin/echo 1-4 > cpus -> set cpus list to cpus 1,2,3,4
583# /bin/echo 1,2,3,4 > cpus -> set cpus list to cpus 1,2,3,4
584
5852.3 Setting flags
586-----------------
587
588The syntax is very simple:
589
590# /bin/echo 1 > cpu_exclusive -> set flag 'cpu_exclusive'
591# /bin/echo 0 > cpu_exclusive -> unset flag 'cpu_exclusive'
592
5932.4 Attaching processes
594-----------------------
595
596# /bin/echo PID > tasks
597
598Note that it is PID, not PIDs. You can only attach ONE task at a time.
599If you have several tasks to attach, you have to do it one after another:
600
601# /bin/echo PID1 > tasks
602# /bin/echo PID2 > tasks
603 ...
604# /bin/echo PIDn > tasks
605
606
6073. Questions
608============
609
610Q: what's up with this '/bin/echo' ?
611A: bash's builtin 'echo' command does not check calls to write() against
612 errors. If you use it in the cpuset file system, you won't be
613 able to tell whether a command succeeded or failed.
614
615Q: When I attach processes, only the first of the line gets really attached !
616A: We can only return one error code per call to write(). So you should also
617 put only ONE pid.
618
6194. Contact
620==========
621
622Web: http://www.bullopensource.org/cpuset