blob: 2f8f24eaefd9ac746f8a1e98b6e2c99abcc23b58 [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
7Portions Copyright (c) 2004 Silicon Graphics, Inc.
8Modified by Paul Jackson <pj@sgi.com>
9
10CONTENTS:
11=========
12
131. Cpusets
14 1.1 What are cpusets ?
15 1.2 Why are cpusets needed ?
16 1.3 How are cpusets implemented ?
17 1.4 How do I use cpusets ?
182. Usage Examples and Syntax
19 2.1 Basic Usage
20 2.2 Adding/removing cpus
21 2.3 Setting flags
22 2.4 Attaching processes
233. Questions
244. Contact
25
261. Cpusets
27==========
28
291.1 What are cpusets ?
30----------------------
31
32Cpusets provide a mechanism for assigning a set of CPUs and Memory
33Nodes to a set of tasks.
34
35Cpusets constrain the CPU and Memory placement of tasks to only
36the resources within a tasks current cpuset. They form a nested
37hierarchy visible in a virtual file system. These are the essential
38hooks, beyond what is already present, required to manage dynamic
39job placement on large systems.
40
41Each task has a pointer to a cpuset. Multiple tasks may reference
42the same cpuset. Requests by a task, using the sched_setaffinity(2)
43system call to include CPUs in its CPU affinity mask, and using the
44mbind(2) and set_mempolicy(2) system calls to include Memory Nodes
45in its memory policy, are both filtered through that tasks cpuset,
46filtering out any CPUs or Memory Nodes not in that cpuset. The
47scheduler will not schedule a task on a CPU that is not allowed in
48its cpus_allowed vector, and the kernel page allocator will not
49allocate a page on a node that is not allowed in the requesting tasks
50mems_allowed vector.
51
52If a cpuset is cpu or mem exclusive, no other cpuset, other than a direct
53ancestor or descendent, may share any of the same CPUs or Memory Nodes.
54
55User level code may create and destroy cpusets by name in the cpuset
56virtual file system, manage the attributes and permissions of these
57cpusets and which CPUs and Memory Nodes are assigned to each cpuset,
58specify and query to which cpuset a task is assigned, and list the
59task pids assigned to a cpuset.
60
61
621.2 Why are cpusets needed ?
63----------------------------
64
65The management of large computer systems, with many processors (CPUs),
66complex memory cache hierarchies and multiple Memory Nodes having
67non-uniform access times (NUMA) presents additional challenges for
68the efficient scheduling and memory placement of processes.
69
70Frequently more modest sized systems can be operated with adequate
71efficiency just by letting the operating system automatically share
72the available CPU and Memory resources amongst the requesting tasks.
73
74But larger systems, which benefit more from careful processor and
75memory placement to reduce memory access times and contention,
76and which typically represent a larger investment for the customer,
77can benefit from explictly placing jobs on properly sized subsets of
78the system.
79
80This can be especially valuable on:
81
82 * Web Servers running multiple instances of the same web application,
83 * Servers running different applications (for instance, a web server
84 and a database), or
85 * NUMA systems running large HPC applications with demanding
86 performance characteristics.
87
88These subsets, or "soft partitions" must be able to be dynamically
89adjusted, as the job mix changes, without impacting other concurrently
90executing jobs.
91
92The kernel cpuset patch provides the minimum essential kernel
93mechanisms required to efficiently implement such subsets. It
94leverages existing CPU and Memory Placement facilities in the Linux
95kernel to avoid any additional impact on the critical scheduler or
96memory allocator code.
97
98
991.3 How are cpusets implemented ?
100---------------------------------
101
102Cpusets provide a Linux kernel (2.6.7 and above) mechanism to constrain
103which CPUs and Memory Nodes are used by a process or set of processes.
104
105The Linux kernel already has a pair of mechanisms to specify on which
106CPUs a task may be scheduled (sched_setaffinity) and on which Memory
107Nodes it may obtain memory (mbind, set_mempolicy).
108
109Cpusets extends these two mechanisms as follows:
110
111 - Cpusets are sets of allowed CPUs and Memory Nodes, known to the
112 kernel.
113 - Each task in the system is attached to a cpuset, via a pointer
114 in the task structure to a reference counted cpuset structure.
115 - Calls to sched_setaffinity are filtered to just those CPUs
116 allowed in that tasks cpuset.
117 - Calls to mbind and set_mempolicy are filtered to just
118 those Memory Nodes allowed in that tasks cpuset.
119 - The root cpuset contains all the systems CPUs and Memory
120 Nodes.
121 - For any cpuset, one can define child cpusets containing a subset
122 of the parents CPU and Memory Node resources.
123 - The hierarchy of cpusets can be mounted at /dev/cpuset, for
124 browsing and manipulation from user space.
125 - A cpuset may be marked exclusive, which ensures that no other
126 cpuset (except direct ancestors and descendents) may contain
127 any overlapping CPUs or Memory Nodes.
128 - You can list all the tasks (by pid) attached to any cpuset.
129
130The implementation of cpusets requires a few, simple hooks
131into the rest of the kernel, none in performance critical paths:
132
133 - in main/init.c, to initialize the root cpuset at system boot.
134 - in fork and exit, to attach and detach a task from its cpuset.
135 - in sched_setaffinity, to mask the requested CPUs by what's
136 allowed in that tasks cpuset.
137 - in sched.c migrate_all_tasks(), to keep migrating tasks within
138 the CPUs allowed by their cpuset, if possible.
139 - in the mbind and set_mempolicy system calls, to mask the requested
140 Memory Nodes by what's allowed in that tasks cpuset.
141 - in page_alloc, to restrict memory to allowed nodes.
142 - in vmscan.c, to restrict page recovery to the current cpuset.
143
144In addition a new file system, of type "cpuset" may be mounted,
145typically at /dev/cpuset, to enable browsing and modifying the cpusets
146presently known to the kernel. No new system calls are added for
147cpusets - all support for querying and modifying cpusets is via
148this cpuset file system.
149
150Each task under /proc has an added file named 'cpuset', displaying
151the cpuset name, as the path relative to the root of the cpuset file
152system.
153
154The /proc/<pid>/status file for each task has two added lines,
155displaying the tasks cpus_allowed (on which CPUs it may be scheduled)
156and mems_allowed (on which Memory Nodes it may obtain memory),
157in the format seen in the following example:
158
159 Cpus_allowed: ffffffff,ffffffff,ffffffff,ffffffff
160 Mems_allowed: ffffffff,ffffffff
161
162Each cpuset is represented by a directory in the cpuset file system
163containing the following files describing that cpuset:
164
165 - cpus: list of CPUs in that cpuset
166 - mems: list of Memory Nodes in that cpuset
167 - cpu_exclusive flag: is cpu placement exclusive?
168 - mem_exclusive flag: is memory placement exclusive?
169 - tasks: list of tasks (by pid) attached to that cpuset
170
171New cpusets are created using the mkdir system call or shell
172command. The properties of a cpuset, such as its flags, allowed
173CPUs and Memory Nodes, and attached tasks, are modified by writing
174to the appropriate file in that cpusets directory, as listed above.
175
176The named hierarchical structure of nested cpusets allows partitioning
177a large system into nested, dynamically changeable, "soft-partitions".
178
179The attachment of each task, automatically inherited at fork by any
180children of that task, to a cpuset allows organizing the work load
181on a system into related sets of tasks such that each set is constrained
182to using the CPUs and Memory Nodes of a particular cpuset. A task
183may be re-attached to any other cpuset, if allowed by the permissions
184on the necessary cpuset file system directories.
185
186Such management of a system "in the large" integrates smoothly with
187the detailed placement done on individual tasks and memory regions
188using the sched_setaffinity, mbind and set_mempolicy system calls.
189
190The following rules apply to each cpuset:
191
192 - Its CPUs and Memory Nodes must be a subset of its parents.
193 - It can only be marked exclusive if its parent is.
194 - If its cpu or memory is exclusive, they may not overlap any sibling.
195
196These rules, and the natural hierarchy of cpusets, enable efficient
197enforcement of the exclusive guarantee, without having to scan all
198cpusets every time any of them change to ensure nothing overlaps a
199exclusive cpuset. Also, the use of a Linux virtual file system (vfs)
200to represent the cpuset hierarchy provides for a familiar permission
201and name space for cpusets, with a minimum of additional kernel code.
202
2031.4 How do I use cpusets ?
204--------------------------
205
206In order to minimize the impact of cpusets on critical kernel
207code, such as the scheduler, and due to the fact that the kernel
208does not support one task updating the memory placement of another
209task directly, the impact on a task of changing its cpuset CPU
210or Memory Node placement, or of changing to which cpuset a task
211is attached, is subtle.
212
213If a cpuset has its Memory Nodes modified, then for each task attached
214to that cpuset, the next time that the kernel attempts to allocate
215a page of memory for that task, the kernel will notice the change
216in the tasks cpuset, and update its per-task memory placement to
217remain within the new cpusets memory placement. If the task was using
218mempolicy MPOL_BIND, and the nodes to which it was bound overlap with
219its new cpuset, then the task will continue to use whatever subset
220of MPOL_BIND nodes are still allowed in the new cpuset. If the task
221was using MPOL_BIND and now none of its MPOL_BIND nodes are allowed
222in the new cpuset, then the task will be essentially treated as if it
223was MPOL_BIND bound to the new cpuset (even though its numa placement,
224as queried by get_mempolicy(), doesn't change). If a task is moved
225from one cpuset to another, then the kernel will adjust the tasks
226memory placement, as above, the next time that the kernel attempts
227to allocate a page of memory for that task.
228
229If a cpuset has its CPUs modified, then each task using that
230cpuset does _not_ change its behavior automatically. In order to
231minimize the impact on the critical scheduling code in the kernel,
232tasks will continue to use their prior CPU placement until they
233are rebound to their cpuset, by rewriting their pid to the 'tasks'
234file of their cpuset. If a task had been bound to some subset of its
235cpuset using the sched_setaffinity() call, and if any of that subset
236is still allowed in its new cpuset settings, then the task will be
237restricted to the intersection of the CPUs it was allowed on before,
238and its new cpuset CPU placement. If, on the other hand, there is
239no overlap between a tasks prior placement and its new cpuset CPU
240placement, then the task will be allowed to run on any CPU allowed
241in its new cpuset. If a task is moved from one cpuset to another,
242its CPU placement is updated in the same way as if the tasks pid is
243rewritten to the 'tasks' file of its current cpuset.
244
245In summary, the memory placement of a task whose cpuset is changed is
246updated by the kernel, on the next allocation of a page for that task,
247but the processor placement is not updated, until that tasks pid is
248rewritten to the 'tasks' file of its cpuset. This is done to avoid
249impacting the scheduler code in the kernel with a check for changes
250in a tasks processor placement.
251
252There is an exception to the above. If hotplug funtionality is used
253to remove all the CPUs that are currently assigned to a cpuset,
254then the kernel will automatically update the cpus_allowed of all
Paul Jacksonb39c4fa2005-05-20 13:59:15 -0700255tasks attached to CPUs in that cpuset to allow all CPUs. When memory
Linus Torvalds1da177e2005-04-16 15:20:36 -0700256hotplug functionality for removing Memory Nodes is available, a
257similar exception is expected to apply there as well. In general,
258the kernel prefers to violate cpuset placement, over starving a task
259that has had all its allowed CPUs or Memory Nodes taken offline. User
260code should reconfigure cpusets to only refer to online CPUs and Memory
261Nodes when using hotplug to add or remove such resources.
262
263There is a second exception to the above. GFP_ATOMIC requests are
264kernel internal allocations that must be satisfied, immediately.
265The kernel may drop some request, in rare cases even panic, if a
266GFP_ATOMIC alloc fails. If the request cannot be satisfied within
267the current tasks cpuset, then we relax the cpuset, and look for
268memory anywhere we can find it. It's better to violate the cpuset
269than stress the kernel.
270
271To start a new job that is to be contained within a cpuset, the steps are:
272
273 1) mkdir /dev/cpuset
274 2) mount -t cpuset none /dev/cpuset
275 3) Create the new cpuset by doing mkdir's and write's (or echo's) in
276 the /dev/cpuset virtual file system.
277 4) Start a task that will be the "founding father" of the new job.
278 5) Attach that task to the new cpuset by writing its pid to the
279 /dev/cpuset tasks file for that cpuset.
280 6) fork, exec or clone the job tasks from this founding father task.
281
282For example, the following sequence of commands will setup a cpuset
283named "Charlie", containing just CPUs 2 and 3, and Memory Node 1,
284and then start a subshell 'sh' in that cpuset:
285
286 mount -t cpuset none /dev/cpuset
287 cd /dev/cpuset
288 mkdir Charlie
289 cd Charlie
290 /bin/echo 2-3 > cpus
291 /bin/echo 1 > mems
292 /bin/echo $$ > tasks
293 sh
294 # The subshell 'sh' is now running in cpuset Charlie
295 # The next line should display '/Charlie'
296 cat /proc/self/cpuset
297
298In the case that a change of cpuset includes wanting to move already
299allocated memory pages, consider further the work of IWAMOTO
300Toshihiro <iwamoto@valinux.co.jp> for page remapping and memory
301hotremoval, which can be found at:
302
303 http://people.valinux.co.jp/~iwamoto/mh.html
304
305The integration of cpusets with such memory migration is not yet
306available.
307
308In the future, a C library interface to cpusets will likely be
309available. For now, the only way to query or modify cpusets is
310via the cpuset file system, using the various cd, mkdir, echo, cat,
311rmdir commands from the shell, or their equivalent from C.
312
313The sched_setaffinity calls can also be done at the shell prompt using
314SGI's runon or Robert Love's taskset. The mbind and set_mempolicy
315calls can be done at the shell prompt using the numactl command
316(part of Andi Kleen's numa package).
317
3182. Usage Examples and Syntax
319============================
320
3212.1 Basic Usage
322---------------
323
324Creating, modifying, using the cpusets can be done through the cpuset
325virtual filesystem.
326
327To mount it, type:
328# mount -t cpuset none /dev/cpuset
329
330Then under /dev/cpuset you can find a tree that corresponds to the
331tree of the cpusets in the system. For instance, /dev/cpuset
332is the cpuset that holds the whole system.
333
334If you want to create a new cpuset under /dev/cpuset:
335# cd /dev/cpuset
336# mkdir my_cpuset
337
338Now you want to do something with this cpuset.
339# cd my_cpuset
340
341In this directory you can find several files:
342# ls
343cpus cpu_exclusive mems mem_exclusive tasks
344
345Reading them will give you information about the state of this cpuset:
346the CPUs and Memory Nodes it can use, the processes that are using
347it, its properties. By writing to these files you can manipulate
348the cpuset.
349
350Set some flags:
351# /bin/echo 1 > cpu_exclusive
352
353Add some cpus:
354# /bin/echo 0-7 > cpus
355
356Now attach your shell to this cpuset:
357# /bin/echo $$ > tasks
358
359You can also create cpusets inside your cpuset by using mkdir in this
360directory.
361# mkdir my_sub_cs
362
363To remove a cpuset, just use rmdir:
364# rmdir my_sub_cs
365This will fail if the cpuset is in use (has cpusets inside, or has
366processes attached).
367
3682.2 Adding/removing cpus
369------------------------
370
371This is the syntax to use when writing in the cpus or mems files
372in cpuset directories:
373
374# /bin/echo 1-4 > cpus -> set cpus list to cpus 1,2,3,4
375# /bin/echo 1,2,3,4 > cpus -> set cpus list to cpus 1,2,3,4
376
3772.3 Setting flags
378-----------------
379
380The syntax is very simple:
381
382# /bin/echo 1 > cpu_exclusive -> set flag 'cpu_exclusive'
383# /bin/echo 0 > cpu_exclusive -> unset flag 'cpu_exclusive'
384
3852.4 Attaching processes
386-----------------------
387
388# /bin/echo PID > tasks
389
390Note that it is PID, not PIDs. You can only attach ONE task at a time.
391If you have several tasks to attach, you have to do it one after another:
392
393# /bin/echo PID1 > tasks
394# /bin/echo PID2 > tasks
395 ...
396# /bin/echo PIDn > tasks
397
398
3993. Questions
400============
401
402Q: what's up with this '/bin/echo' ?
403A: bash's builtin 'echo' command does not check calls to write() against
404 errors. If you use it in the cpuset file system, you won't be
405 able to tell whether a command succeeded or failed.
406
407Q: When I attach processes, only the first of the line gets really attached !
408A: We can only return one error code per call to write(). So you should also
409 put only ONE pid.
410
4114. Contact
412==========
413
414Web: http://www.bullopensource.org/cpuset