Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | CPUSETS |
| 2 | ------- |
| 3 | |
| 4 | Copyright (C) 2004 BULL SA. |
| 5 | Written by Simon.Derr@bull.net |
| 6 | |
| 7 | Portions Copyright (c) 2004 Silicon Graphics, Inc. |
| 8 | Modified by Paul Jackson <pj@sgi.com> |
| 9 | |
| 10 | CONTENTS: |
| 11 | ========= |
| 12 | |
| 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 How do I use cpusets ? |
| 18 | 2. 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 |
| 23 | 3. Questions |
| 24 | 4. Contact |
| 25 | |
| 26 | 1. Cpusets |
| 27 | ========== |
| 28 | |
| 29 | 1.1 What are cpusets ? |
| 30 | ---------------------- |
| 31 | |
| 32 | Cpusets provide a mechanism for assigning a set of CPUs and Memory |
| 33 | Nodes to a set of tasks. |
| 34 | |
| 35 | Cpusets constrain the CPU and Memory placement of tasks to only |
| 36 | the resources within a tasks current cpuset. They form a nested |
| 37 | hierarchy visible in a virtual file system. These are the essential |
| 38 | hooks, beyond what is already present, required to manage dynamic |
| 39 | job placement on large systems. |
| 40 | |
| 41 | Each task has a pointer to a cpuset. Multiple tasks may reference |
| 42 | the same cpuset. Requests by a task, using the sched_setaffinity(2) |
| 43 | system call to include CPUs in its CPU affinity mask, and using the |
| 44 | mbind(2) and set_mempolicy(2) system calls to include Memory Nodes |
| 45 | in its memory policy, are both filtered through that tasks cpuset, |
| 46 | filtering out any CPUs or Memory Nodes not in that cpuset. The |
| 47 | scheduler will not schedule a task on a CPU that is not allowed in |
| 48 | its cpus_allowed vector, and the kernel page allocator will not |
| 49 | allocate a page on a node that is not allowed in the requesting tasks |
| 50 | mems_allowed vector. |
| 51 | |
| 52 | If a cpuset is cpu or mem exclusive, no other cpuset, other than a direct |
| 53 | ancestor or descendent, may share any of the same CPUs or Memory Nodes. |
Dinakar Guniguntala | 85d7b94 | 2005-06-25 14:57:34 -0700 | [diff] [blame] | 54 | A cpuset that is cpu exclusive has a sched domain associated with it. |
| 55 | The sched domain consists of all cpus in the current cpuset that are not |
| 56 | part of any exclusive child cpusets. |
| 57 | This ensures that the scheduler load balacing code only balances |
| 58 | against the cpus that are in the sched domain as defined above and not |
| 59 | all of the cpus in the system. This removes any overhead due to |
| 60 | load balancing code trying to pull tasks outside of the cpu exclusive |
| 61 | cpuset only to be prevented by the tasks' cpus_allowed mask. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 62 | |
Paul Jackson | 9bf2229 | 2005-09-06 15:18:12 -0700 | [diff] [blame] | 63 | A cpuset that is mem_exclusive restricts kernel allocations for |
| 64 | page, buffer and other data commonly shared by the kernel across |
| 65 | multiple users. All cpusets, whether mem_exclusive or not, restrict |
| 66 | allocations of memory for user space. This enables configuring a |
| 67 | system so that several independent jobs can share common kernel |
| 68 | data, such as file system pages, while isolating each jobs user |
| 69 | allocation in its own cpuset. To do this, construct a large |
| 70 | mem_exclusive cpuset to hold all the jobs, and construct child, |
| 71 | non-mem_exclusive cpusets for each individual job. Only a small |
| 72 | amount of typical kernel memory, such as requests from interrupt |
| 73 | handlers, is allowed to be taken outside even a mem_exclusive cpuset. |
| 74 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 75 | User level code may create and destroy cpusets by name in the cpuset |
| 76 | virtual file system, manage the attributes and permissions of these |
| 77 | cpusets and which CPUs and Memory Nodes are assigned to each cpuset, |
| 78 | specify and query to which cpuset a task is assigned, and list the |
| 79 | task pids assigned to a cpuset. |
| 80 | |
| 81 | |
| 82 | 1.2 Why are cpusets needed ? |
| 83 | ---------------------------- |
| 84 | |
| 85 | The management of large computer systems, with many processors (CPUs), |
| 86 | complex memory cache hierarchies and multiple Memory Nodes having |
| 87 | non-uniform access times (NUMA) presents additional challenges for |
| 88 | the efficient scheduling and memory placement of processes. |
| 89 | |
| 90 | Frequently more modest sized systems can be operated with adequate |
| 91 | efficiency just by letting the operating system automatically share |
| 92 | the available CPU and Memory resources amongst the requesting tasks. |
| 93 | |
| 94 | But larger systems, which benefit more from careful processor and |
| 95 | memory placement to reduce memory access times and contention, |
| 96 | and which typically represent a larger investment for the customer, |
| 97 | can benefit from explictly placing jobs on properly sized subsets of |
| 98 | the system. |
| 99 | |
| 100 | This can be especially valuable on: |
| 101 | |
| 102 | * Web Servers running multiple instances of the same web application, |
| 103 | * Servers running different applications (for instance, a web server |
| 104 | and a database), or |
| 105 | * NUMA systems running large HPC applications with demanding |
| 106 | performance characteristics. |
Dinakar Guniguntala | 85d7b94 | 2005-06-25 14:57:34 -0700 | [diff] [blame] | 107 | * Also cpu_exclusive cpusets are useful for servers running orthogonal |
| 108 | workloads such as RT applications requiring low latency and HPC |
| 109 | applications that are throughput sensitive |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 110 | |
| 111 | These subsets, or "soft partitions" must be able to be dynamically |
| 112 | adjusted, as the job mix changes, without impacting other concurrently |
| 113 | executing jobs. |
| 114 | |
| 115 | The kernel cpuset patch provides the minimum essential kernel |
| 116 | mechanisms required to efficiently implement such subsets. It |
| 117 | leverages existing CPU and Memory Placement facilities in the Linux |
| 118 | kernel to avoid any additional impact on the critical scheduler or |
| 119 | memory allocator code. |
| 120 | |
| 121 | |
| 122 | 1.3 How are cpusets implemented ? |
| 123 | --------------------------------- |
| 124 | |
| 125 | Cpusets provide a Linux kernel (2.6.7 and above) mechanism to constrain |
| 126 | which CPUs and Memory Nodes are used by a process or set of processes. |
| 127 | |
| 128 | The Linux kernel already has a pair of mechanisms to specify on which |
| 129 | CPUs a task may be scheduled (sched_setaffinity) and on which Memory |
| 130 | Nodes it may obtain memory (mbind, set_mempolicy). |
| 131 | |
| 132 | Cpusets extends these two mechanisms as follows: |
| 133 | |
| 134 | - Cpusets are sets of allowed CPUs and Memory Nodes, known to the |
| 135 | kernel. |
| 136 | - Each task in the system is attached to a cpuset, via a pointer |
| 137 | in the task structure to a reference counted cpuset structure. |
| 138 | - Calls to sched_setaffinity are filtered to just those CPUs |
| 139 | allowed in that tasks cpuset. |
| 140 | - Calls to mbind and set_mempolicy are filtered to just |
| 141 | those Memory Nodes allowed in that tasks cpuset. |
| 142 | - The root cpuset contains all the systems CPUs and Memory |
| 143 | Nodes. |
| 144 | - For any cpuset, one can define child cpusets containing a subset |
| 145 | of the parents CPU and Memory Node resources. |
| 146 | - The hierarchy of cpusets can be mounted at /dev/cpuset, for |
| 147 | browsing and manipulation from user space. |
| 148 | - A cpuset may be marked exclusive, which ensures that no other |
| 149 | cpuset (except direct ancestors and descendents) may contain |
| 150 | any overlapping CPUs or Memory Nodes. |
Dinakar Guniguntala | 85d7b94 | 2005-06-25 14:57:34 -0700 | [diff] [blame] | 151 | Also a cpu_exclusive cpuset would be associated with a sched |
| 152 | domain. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 153 | - You can list all the tasks (by pid) attached to any cpuset. |
| 154 | |
| 155 | The implementation of cpusets requires a few, simple hooks |
| 156 | into the rest of the kernel, none in performance critical paths: |
| 157 | |
| 158 | - in main/init.c, to initialize the root cpuset at system boot. |
| 159 | - in fork and exit, to attach and detach a task from its cpuset. |
| 160 | - in sched_setaffinity, to mask the requested CPUs by what's |
| 161 | allowed in that tasks cpuset. |
| 162 | - in sched.c migrate_all_tasks(), to keep migrating tasks within |
| 163 | the CPUs allowed by their cpuset, if possible. |
Dinakar Guniguntala | 85d7b94 | 2005-06-25 14:57:34 -0700 | [diff] [blame] | 164 | - in sched.c, a new API partition_sched_domains for handling |
| 165 | sched domain changes associated with cpu_exclusive cpusets |
| 166 | and related changes in both sched.c and arch/ia64/kernel/domain.c |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 167 | - in the mbind and set_mempolicy system calls, to mask the requested |
| 168 | Memory Nodes by what's allowed in that tasks cpuset. |
| 169 | - in page_alloc, to restrict memory to allowed nodes. |
| 170 | - in vmscan.c, to restrict page recovery to the current cpuset. |
| 171 | |
| 172 | In addition a new file system, of type "cpuset" may be mounted, |
| 173 | typically at /dev/cpuset, to enable browsing and modifying the cpusets |
| 174 | presently known to the kernel. No new system calls are added for |
| 175 | cpusets - all support for querying and modifying cpusets is via |
| 176 | this cpuset file system. |
| 177 | |
| 178 | Each task under /proc has an added file named 'cpuset', displaying |
| 179 | the cpuset name, as the path relative to the root of the cpuset file |
| 180 | system. |
| 181 | |
| 182 | The /proc/<pid>/status file for each task has two added lines, |
| 183 | displaying the tasks cpus_allowed (on which CPUs it may be scheduled) |
| 184 | and mems_allowed (on which Memory Nodes it may obtain memory), |
| 185 | in the format seen in the following example: |
| 186 | |
| 187 | Cpus_allowed: ffffffff,ffffffff,ffffffff,ffffffff |
| 188 | Mems_allowed: ffffffff,ffffffff |
| 189 | |
| 190 | Each cpuset is represented by a directory in the cpuset file system |
| 191 | containing the following files describing that cpuset: |
| 192 | |
| 193 | - cpus: list of CPUs in that cpuset |
| 194 | - mems: list of Memory Nodes in that cpuset |
| 195 | - cpu_exclusive flag: is cpu placement exclusive? |
| 196 | - mem_exclusive flag: is memory placement exclusive? |
| 197 | - tasks: list of tasks (by pid) attached to that cpuset |
| 198 | |
| 199 | New cpusets are created using the mkdir system call or shell |
| 200 | command. The properties of a cpuset, such as its flags, allowed |
| 201 | CPUs and Memory Nodes, and attached tasks, are modified by writing |
| 202 | to the appropriate file in that cpusets directory, as listed above. |
| 203 | |
| 204 | The named hierarchical structure of nested cpusets allows partitioning |
| 205 | a large system into nested, dynamically changeable, "soft-partitions". |
| 206 | |
| 207 | The attachment of each task, automatically inherited at fork by any |
| 208 | children of that task, to a cpuset allows organizing the work load |
| 209 | on a system into related sets of tasks such that each set is constrained |
| 210 | to using the CPUs and Memory Nodes of a particular cpuset. A task |
| 211 | may be re-attached to any other cpuset, if allowed by the permissions |
| 212 | on the necessary cpuset file system directories. |
| 213 | |
| 214 | Such management of a system "in the large" integrates smoothly with |
| 215 | the detailed placement done on individual tasks and memory regions |
| 216 | using the sched_setaffinity, mbind and set_mempolicy system calls. |
| 217 | |
| 218 | The following rules apply to each cpuset: |
| 219 | |
| 220 | - Its CPUs and Memory Nodes must be a subset of its parents. |
| 221 | - It can only be marked exclusive if its parent is. |
| 222 | - If its cpu or memory is exclusive, they may not overlap any sibling. |
| 223 | |
| 224 | These rules, and the natural hierarchy of cpusets, enable efficient |
| 225 | enforcement of the exclusive guarantee, without having to scan all |
| 226 | cpusets every time any of them change to ensure nothing overlaps a |
| 227 | exclusive cpuset. Also, the use of a Linux virtual file system (vfs) |
| 228 | to represent the cpuset hierarchy provides for a familiar permission |
| 229 | and name space for cpusets, with a minimum of additional kernel code. |
| 230 | |
| 231 | 1.4 How do I use cpusets ? |
| 232 | -------------------------- |
| 233 | |
| 234 | In order to minimize the impact of cpusets on critical kernel |
| 235 | code, such as the scheduler, and due to the fact that the kernel |
| 236 | does not support one task updating the memory placement of another |
| 237 | task directly, the impact on a task of changing its cpuset CPU |
| 238 | or Memory Node placement, or of changing to which cpuset a task |
| 239 | is attached, is subtle. |
| 240 | |
| 241 | If a cpuset has its Memory Nodes modified, then for each task attached |
| 242 | to that cpuset, the next time that the kernel attempts to allocate |
| 243 | a page of memory for that task, the kernel will notice the change |
| 244 | in the tasks cpuset, and update its per-task memory placement to |
| 245 | remain within the new cpusets memory placement. If the task was using |
| 246 | mempolicy MPOL_BIND, and the nodes to which it was bound overlap with |
| 247 | its new cpuset, then the task will continue to use whatever subset |
| 248 | of MPOL_BIND nodes are still allowed in the new cpuset. If the task |
| 249 | was using MPOL_BIND and now none of its MPOL_BIND nodes are allowed |
| 250 | in the new cpuset, then the task will be essentially treated as if it |
| 251 | was MPOL_BIND bound to the new cpuset (even though its numa placement, |
| 252 | as queried by get_mempolicy(), doesn't change). If a task is moved |
| 253 | from one cpuset to another, then the kernel will adjust the tasks |
| 254 | memory placement, as above, the next time that the kernel attempts |
| 255 | to allocate a page of memory for that task. |
| 256 | |
| 257 | If a cpuset has its CPUs modified, then each task using that |
| 258 | cpuset does _not_ change its behavior automatically. In order to |
| 259 | minimize the impact on the critical scheduling code in the kernel, |
| 260 | tasks will continue to use their prior CPU placement until they |
| 261 | are rebound to their cpuset, by rewriting their pid to the 'tasks' |
| 262 | file of their cpuset. If a task had been bound to some subset of its |
| 263 | cpuset using the sched_setaffinity() call, and if any of that subset |
| 264 | is still allowed in its new cpuset settings, then the task will be |
| 265 | restricted to the intersection of the CPUs it was allowed on before, |
| 266 | and its new cpuset CPU placement. If, on the other hand, there is |
| 267 | no overlap between a tasks prior placement and its new cpuset CPU |
| 268 | placement, then the task will be allowed to run on any CPU allowed |
| 269 | in its new cpuset. If a task is moved from one cpuset to another, |
| 270 | its CPU placement is updated in the same way as if the tasks pid is |
| 271 | rewritten to the 'tasks' file of its current cpuset. |
| 272 | |
| 273 | In summary, the memory placement of a task whose cpuset is changed is |
| 274 | updated by the kernel, on the next allocation of a page for that task, |
| 275 | but the processor placement is not updated, until that tasks pid is |
| 276 | rewritten to the 'tasks' file of its cpuset. This is done to avoid |
| 277 | impacting the scheduler code in the kernel with a check for changes |
| 278 | in a tasks processor placement. |
| 279 | |
| 280 | There is an exception to the above. If hotplug funtionality is used |
| 281 | to remove all the CPUs that are currently assigned to a cpuset, |
| 282 | then the kernel will automatically update the cpus_allowed of all |
Paul Jackson | b39c4fa | 2005-05-20 13:59:15 -0700 | [diff] [blame] | 283 | tasks attached to CPUs in that cpuset to allow all CPUs. When memory |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 284 | hotplug functionality for removing Memory Nodes is available, a |
| 285 | similar exception is expected to apply there as well. In general, |
| 286 | the kernel prefers to violate cpuset placement, over starving a task |
| 287 | that has had all its allowed CPUs or Memory Nodes taken offline. User |
| 288 | code should reconfigure cpusets to only refer to online CPUs and Memory |
| 289 | Nodes when using hotplug to add or remove such resources. |
| 290 | |
| 291 | There is a second exception to the above. GFP_ATOMIC requests are |
| 292 | kernel internal allocations that must be satisfied, immediately. |
| 293 | The kernel may drop some request, in rare cases even panic, if a |
| 294 | GFP_ATOMIC alloc fails. If the request cannot be satisfied within |
| 295 | the current tasks cpuset, then we relax the cpuset, and look for |
| 296 | memory anywhere we can find it. It's better to violate the cpuset |
| 297 | than stress the kernel. |
| 298 | |
| 299 | To start a new job that is to be contained within a cpuset, the steps are: |
| 300 | |
| 301 | 1) mkdir /dev/cpuset |
| 302 | 2) mount -t cpuset none /dev/cpuset |
| 303 | 3) Create the new cpuset by doing mkdir's and write's (or echo's) in |
| 304 | the /dev/cpuset virtual file system. |
| 305 | 4) Start a task that will be the "founding father" of the new job. |
| 306 | 5) Attach that task to the new cpuset by writing its pid to the |
| 307 | /dev/cpuset tasks file for that cpuset. |
| 308 | 6) fork, exec or clone the job tasks from this founding father task. |
| 309 | |
| 310 | For example, the following sequence of commands will setup a cpuset |
| 311 | named "Charlie", containing just CPUs 2 and 3, and Memory Node 1, |
| 312 | and then start a subshell 'sh' in that cpuset: |
| 313 | |
| 314 | mount -t cpuset none /dev/cpuset |
| 315 | cd /dev/cpuset |
| 316 | mkdir Charlie |
| 317 | cd Charlie |
| 318 | /bin/echo 2-3 > cpus |
| 319 | /bin/echo 1 > mems |
| 320 | /bin/echo $$ > tasks |
| 321 | sh |
| 322 | # The subshell 'sh' is now running in cpuset Charlie |
| 323 | # The next line should display '/Charlie' |
| 324 | cat /proc/self/cpuset |
| 325 | |
| 326 | In the case that a change of cpuset includes wanting to move already |
| 327 | allocated memory pages, consider further the work of IWAMOTO |
| 328 | Toshihiro <iwamoto@valinux.co.jp> for page remapping and memory |
| 329 | hotremoval, which can be found at: |
| 330 | |
| 331 | http://people.valinux.co.jp/~iwamoto/mh.html |
| 332 | |
| 333 | The integration of cpusets with such memory migration is not yet |
| 334 | available. |
| 335 | |
| 336 | In the future, a C library interface to cpusets will likely be |
| 337 | available. For now, the only way to query or modify cpusets is |
| 338 | via the cpuset file system, using the various cd, mkdir, echo, cat, |
| 339 | rmdir commands from the shell, or their equivalent from C. |
| 340 | |
| 341 | The sched_setaffinity calls can also be done at the shell prompt using |
| 342 | SGI's runon or Robert Love's taskset. The mbind and set_mempolicy |
| 343 | calls can be done at the shell prompt using the numactl command |
| 344 | (part of Andi Kleen's numa package). |
| 345 | |
| 346 | 2. Usage Examples and Syntax |
| 347 | ============================ |
| 348 | |
| 349 | 2.1 Basic Usage |
| 350 | --------------- |
| 351 | |
| 352 | Creating, modifying, using the cpusets can be done through the cpuset |
| 353 | virtual filesystem. |
| 354 | |
| 355 | To mount it, type: |
| 356 | # mount -t cpuset none /dev/cpuset |
| 357 | |
| 358 | Then under /dev/cpuset you can find a tree that corresponds to the |
| 359 | tree of the cpusets in the system. For instance, /dev/cpuset |
| 360 | is the cpuset that holds the whole system. |
| 361 | |
| 362 | If you want to create a new cpuset under /dev/cpuset: |
| 363 | # cd /dev/cpuset |
| 364 | # mkdir my_cpuset |
| 365 | |
| 366 | Now you want to do something with this cpuset. |
| 367 | # cd my_cpuset |
| 368 | |
| 369 | In this directory you can find several files: |
| 370 | # ls |
| 371 | cpus cpu_exclusive mems mem_exclusive tasks |
| 372 | |
| 373 | Reading them will give you information about the state of this cpuset: |
| 374 | the CPUs and Memory Nodes it can use, the processes that are using |
| 375 | it, its properties. By writing to these files you can manipulate |
| 376 | the cpuset. |
| 377 | |
| 378 | Set some flags: |
| 379 | # /bin/echo 1 > cpu_exclusive |
| 380 | |
| 381 | Add some cpus: |
| 382 | # /bin/echo 0-7 > cpus |
| 383 | |
| 384 | Now attach your shell to this cpuset: |
| 385 | # /bin/echo $$ > tasks |
| 386 | |
| 387 | You can also create cpusets inside your cpuset by using mkdir in this |
| 388 | directory. |
| 389 | # mkdir my_sub_cs |
| 390 | |
| 391 | To remove a cpuset, just use rmdir: |
| 392 | # rmdir my_sub_cs |
| 393 | This will fail if the cpuset is in use (has cpusets inside, or has |
| 394 | processes attached). |
| 395 | |
| 396 | 2.2 Adding/removing cpus |
| 397 | ------------------------ |
| 398 | |
| 399 | This is the syntax to use when writing in the cpus or mems files |
| 400 | in cpuset directories: |
| 401 | |
| 402 | # /bin/echo 1-4 > cpus -> set cpus list to cpus 1,2,3,4 |
| 403 | # /bin/echo 1,2,3,4 > cpus -> set cpus list to cpus 1,2,3,4 |
| 404 | |
| 405 | 2.3 Setting flags |
| 406 | ----------------- |
| 407 | |
| 408 | The syntax is very simple: |
| 409 | |
| 410 | # /bin/echo 1 > cpu_exclusive -> set flag 'cpu_exclusive' |
| 411 | # /bin/echo 0 > cpu_exclusive -> unset flag 'cpu_exclusive' |
| 412 | |
| 413 | 2.4 Attaching processes |
| 414 | ----------------------- |
| 415 | |
| 416 | # /bin/echo PID > tasks |
| 417 | |
| 418 | Note that it is PID, not PIDs. You can only attach ONE task at a time. |
| 419 | If you have several tasks to attach, you have to do it one after another: |
| 420 | |
| 421 | # /bin/echo PID1 > tasks |
| 422 | # /bin/echo PID2 > tasks |
| 423 | ... |
| 424 | # /bin/echo PIDn > tasks |
| 425 | |
| 426 | |
| 427 | 3. Questions |
| 428 | ============ |
| 429 | |
| 430 | Q: what's up with this '/bin/echo' ? |
| 431 | A: bash's builtin 'echo' command does not check calls to write() against |
| 432 | errors. If you use it in the cpuset file system, you won't be |
| 433 | able to tell whether a command succeeded or failed. |
| 434 | |
| 435 | Q: When I attach processes, only the first of the line gets really attached ! |
| 436 | A: We can only return one error code per call to write(). So you should also |
| 437 | put only ONE pid. |
| 438 | |
| 439 | 4. Contact |
| 440 | ========== |
| 441 | |
| 442 | Web: http://www.bullopensource.org/cpuset |