1                                CPUSETS
   2                                -------
   4Copyright (C) 2004 BULL SA.
   5Written by
   7Portions Copyright (c) 2004-2006 Silicon Graphics, Inc.
   8Modified by Paul Jackson <>
   9Modified by Christoph Lameter <>
  10Modified by Paul Menage <>
  11Modified by Hidetoshi Seto <>
  161. Cpusets
  17  1.1 What are cpusets ?
  18  1.2 Why are cpusets needed ?
  19  1.3 How are cpusets implemented ?
  20  1.4 What are exclusive cpusets ?
  21  1.5 What is memory_pressure ?
  22  1.6 What is memory spread ?
  23  1.7 What is sched_load_balance ?
  24  1.8 What is sched_relax_domain_level ?
  25  1.9 How do I use cpusets ?
  262. Usage Examples and Syntax
  27  2.1 Basic Usage
  28  2.2 Adding/removing cpus
  29  2.3 Setting flags
  30  2.4 Attaching processes
  313. Questions
  324. Contact
  341. Cpusets
  371.1 What are cpusets ?
  40Cpusets provide a mechanism for assigning a set of CPUs and Memory
  41Nodes to a set of tasks.   In this document "Memory Node" refers to
  42an on-line node that contains memory.
  44Cpusets constrain the CPU and Memory placement of tasks to only
  45the resources within a tasks current cpuset.  They form a nested
  46hierarchy visible in a virtual file system.  These are the essential
  47hooks, beyond what is already present, required to manage dynamic
  48job placement on large systems.
  50Cpusets use the generic cgroup subsystem described in
  53Requests by a task, using the sched_setaffinity(2) system call to
  54include CPUs in its CPU affinity mask, and using the mbind(2) and
  55set_mempolicy(2) system calls to include Memory Nodes in its memory
  56policy, are both filtered through that tasks cpuset, filtering out any
  57CPUs or Memory Nodes not in that cpuset.  The scheduler will not
  58schedule a task on a CPU that is not allowed in its cpus_allowed
  59vector, and the kernel page allocator will not allocate a page on a
  60node that is not allowed in the requesting tasks mems_allowed vector.
  62User level code may create and destroy cpusets by name in the cgroup
  63virtual file system, manage the attributes and permissions of these
  64cpusets and which CPUs and Memory Nodes are assigned to each cpuset,
  65specify and query to which cpuset a task is assigned, and list the
  66task pids assigned to a cpuset.
  691.2 Why are cpusets needed ?
  72The management of large computer systems, with many processors (CPUs),
  73complex memory cache hierarchies and multiple Memory Nodes having
  74non-uniform access times (NUMA) presents additional challenges for
  75the efficient scheduling and memory placement of processes.
  77Frequently more modest sized systems can be operated with adequate
  78efficiency just by letting the operating system automatically share
  79the available CPU and Memory resources amongst the requesting tasks.
  81But larger systems, which benefit more from careful processor and
  82memory placement to reduce memory access times and contention,
  83and which typically represent a larger investment for the customer,
  84can benefit from explicitly placing jobs on properly sized subsets of
  85the system.
  87This can be especially valuable on:
  89    * Web Servers running multiple instances of the same web application,
  90    * Servers running different applications (for instance, a web server
  91      and a database), or
  92    * NUMA systems running large HPC applications with demanding
  93      performance characteristics.
  95These subsets, or "soft partitions" must be able to be dynamically
  96adjusted, as the job mix changes, without impacting other concurrently
  97executing jobs. The location of the running jobs pages may also be moved
  98when the memory locations are changed.
 100The kernel cpuset patch provides the minimum essential kernel
 101mechanisms required to efficiently implement such subsets.  It
 102leverages existing CPU and Memory Placement facilities in the Linux
 103kernel to avoid any additional impact on the critical scheduler or
 104memory allocator code.
 1071.3 How are cpusets implemented ?
 110Cpusets provide a Linux kernel mechanism to constrain which CPUs and
 111Memory Nodes are used by a process or set of processes.
 113The Linux kernel already has a pair of mechanisms to specify on which
 114CPUs a task may be scheduled (sched_setaffinity) and on which Memory
 115Nodes it may obtain memory (mbind, set_mempolicy).
 117Cpusets extends these two mechanisms as follows:
 119 - Cpusets are sets of allowed CPUs and Memory Nodes, known to the
 120   kernel.
 121 - Each task in the system is attached to a cpuset, via a pointer
 122   in the task structure to a reference counted cgroup structure.
 123 - Calls to sched_setaffinity are filtered to just those CPUs
 124   allowed in that tasks cpuset.
 125 - Calls to mbind and set_mempolicy are filtered to just
 126   those Memory Nodes allowed in that tasks cpuset.
 127 - The root cpuset contains all the systems CPUs and Memory
 128   Nodes.
 129 - For any cpuset, one can define child cpusets containing a subset
 130   of the parents CPU and Memory Node resources.
 131 - The hierarchy of cpusets can be mounted at /dev/cpuset, for
 132   browsing and manipulation from user space.
 133 - A cpuset may be marked exclusive, which ensures that no other
 134   cpuset (except direct ancestors and descendents) may contain
 135   any overlapping CPUs or Memory Nodes.
 136 - You can list all the tasks (by pid) attached to any cpuset.
 138The implementation of cpusets requires a few, simple hooks
 139into the rest of the kernel, none in performance critical paths:
 141 - in init/main.c, to initialize the root cpuset at system boot.
 142 - in fork and exit, to attach and detach a task from its cpuset.
 143 - in sched_setaffinity, to mask the requested CPUs by what's
 144   allowed in that tasks cpuset.
 145 - in sched.c migrate_live_tasks(), to keep migrating tasks within
 146   the CPUs allowed by their cpuset, if possible.
 147 - in the mbind and set_mempolicy system calls, to mask the requested
 148   Memory Nodes by what's allowed in that tasks cpuset.
 149 - in page_alloc.c, to restrict memory to allowed nodes.
 150 - in vmscan.c, to restrict page recovery to the current cpuset.
 152You should mount the "cgroup" filesystem type in order to enable
 153browsing and modifying the cpusets presently known to the kernel.  No
 154new system calls are added for cpusets - all support for querying and
 155modifying cpusets is via this cpuset file system.
 157The /proc/<pid>/status file for each task has four added lines,
 158displaying the tasks cpus_allowed (on which CPUs it may be scheduled)
 159and mems_allowed (on which Memory Nodes it may obtain memory),
 160in the two formats seen in the following example:
 162  Cpus_allowed:   ffffffff,ffffffff,ffffffff,ffffffff
 163  Cpus_allowed_list:      0-127
 164  Mems_allowed:   ffffffff,ffffffff
 165  Mems_allowed_list:      0-63
 167Each cpuset is represented by a directory in the cgroup file system
 168containing (on top of the standard cgroup files) the following
 169files describing that cpuset:
 171 - cpus: list of CPUs in that cpuset
 172 - mems: list of Memory Nodes in that cpuset
 173 - memory_migrate flag: if set, move pages to cpusets nodes
 174 - cpu_exclusive flag: is cpu placement exclusive?
 175 - mem_exclusive flag: is memory placement exclusive?
 176 - mem_hardwall flag:  is memory allocation hardwalled
 177 - memory_pressure: measure of how much paging pressure in cpuset
 178 - memory_spread_page flag: if set, spread page cache evenly on allowed nodes
 179 - memory_spread_slab flag: if set, spread slab cache evenly on allowed nodes
 180 - sched_load_balance flag: if set, load balance within CPUs on that cpuset
 181 - sched_relax_domain_level: the searching range when migrating tasks
 183In addition, the root cpuset only has the following file:
 184 - memory_pressure_enabled flag: compute memory_pressure?
 186New cpusets are created using the mkdir system call or shell
 187command.  The properties of a cpuset, such as its flags, allowed
 188CPUs and Memory Nodes, and attached tasks, are modified by writing
 189to the appropriate file in that cpusets directory, as listed above.
 191The named hierarchical structure of nested cpusets allows partitioning
 192a large system into nested, dynamically changeable, "soft-partitions".
 194The attachment of each task, automatically inherited at fork by any
 195children of that task, to a cpuset allows organizing the work load
 196on a system into related sets of tasks such that each set is constrained
 197to using the CPUs and Memory Nodes of a particular cpuset.  A task
 198may be re-attached to any other cpuset, if allowed by the permissions
 199on the necessary cpuset file system directories.
 201Such management of a system "in the large" integrates smoothly with
 202the detailed placement done on individual tasks and memory regions
 203using the sched_setaffinity, mbind and set_mempolicy system calls.
 205The following rules apply to each cpuset:
 207 - Its CPUs and Memory Nodes must be a subset of its parents.
 208 - It can't be marked exclusive unless its parent is.
 209 - If its cpu or memory is exclusive, they may not overlap any sibling.
 211These rules, and the natural hierarchy of cpusets, enable efficient
 212enforcement of the exclusive guarantee, without having to scan all
 213cpusets every time any of them change to ensure nothing overlaps a
 214exclusive cpuset.  Also, the use of a Linux virtual file system (vfs)
 215to represent the cpuset hierarchy provides for a familiar permission
 216and name space for cpusets, with a minimum of additional kernel code.
 218The cpus and mems files in the root (top_cpuset) cpuset are
 219read-only.  The cpus file automatically tracks the value of
 220cpu_online_map using a CPU hotplug notifier, and the mems file
 221automatically tracks the value of node_states[N_HIGH_MEMORY]--i.e.,
 222nodes with memory--using the cpuset_track_online_nodes() hook.
 2251.4 What are exclusive cpusets ?
 228If a cpuset is cpu or mem exclusive, no other cpuset, other than
 229a direct ancestor or descendent, may share any of the same CPUs or
 230Memory Nodes.
 232A cpuset that is mem_exclusive *or* mem_hardwall is "hardwalled",
 233i.e. it restricts kernel allocations for page, buffer and other data
 234commonly shared by the kernel across multiple users.  All cpusets,
 235whether hardwalled or not, restrict allocations of memory for user
 236space.  This enables configuring a system so that several independent
 237jobs can share common kernel data, such as file system pages, while
 238isolating each job's user allocation in its own cpuset.  To do this,
 239construct a large mem_exclusive cpuset to hold all the jobs, and
 240construct child, non-mem_exclusive cpusets for each individual job.
 241Only a small amount of typical kernel memory, such as requests from
 242interrupt handlers, is allowed to be taken outside even a
 243mem_exclusive cpuset.
 2461.5 What is memory_pressure ?
 248The memory_pressure of a cpuset provides a simple per-cpuset metric
 249of the rate that the tasks in a cpuset are attempting to free up in
 250use memory on the nodes of the cpuset to satisfy additional memory
 253This enables batch managers monitoring jobs running in dedicated
 254cpusets to efficiently detect what level of memory pressure that job
 255is causing.
 257This is useful both on tightly managed systems running a wide mix of
 258submitted jobs, which may choose to terminate or re-prioritize jobs that
 259are trying to use more memory than allowed on the nodes assigned to them,
 260and with tightly coupled, long running, massively parallel scientific
 261computing jobs that will dramatically fail to meet required performance
 262goals if they start to use more memory than allowed to them.
 264This mechanism provides a very economical way for the batch manager
 265to monitor a cpuset for signs of memory pressure.  It's up to the
 266batch manager or other user code to decide what to do about it and
 267take action.
 269==> Unless this feature is enabled by writing "1" to the special file
 270    /dev/cpuset/memory_pressure_enabled, the hook in the rebalance
 271    code of __alloc_pages() for this metric reduces to simply noticing
 272    that the cpuset_memory_pressure_enabled flag is zero.  So only
 273    systems that enable this feature will compute the metric.
 275Why a per-cpuset, running average:
 277    Because this meter is per-cpuset, rather than per-task or mm,
 278    the system load imposed by a batch scheduler monitoring this
 279    metric is sharply reduced on large systems, because a scan of
 280    the tasklist can be avoided on each set of queries.
 282    Because this meter is a running average, instead of an accumulating
 283    counter, a batch scheduler can detect memory pressure with a
 284    single read, instead of having to read and accumulate results
 285    for a period of time.
 287    Because this meter is per-cpuset rather than per-task or mm,
 288    the batch scheduler can obtain the key information, memory
 289    pressure in a cpuset, with a single read, rather than having to
 290    query and accumulate results over all the (dynamically changing)
 291    set of tasks in the cpuset.
 293A per-cpuset simple digital filter (requires a spinlock and 3 words
 294of data per-cpuset) is kept, and updated by any task attached to that
 295cpuset, if it enters the synchronous (direct) page reclaim code.
 297A per-cpuset file provides an integer number representing the recent
 298(half-life of 10 seconds) rate of direct page reclaims caused by
 299the tasks in the cpuset, in units of reclaims attempted per second,
 300times 1000.
 3031.6 What is memory spread ?
 305There are two boolean flag files per cpuset that control where the
 306kernel allocates pages for the file system buffers and related in
 307kernel data structures.  They are called 'memory_spread_page' and
 310If the per-cpuset boolean flag file 'memory_spread_page' is set, then
 311the kernel will spread the file system buffers (page cache) evenly
 312over all the nodes that the faulting task is allowed to use, instead
 313of preferring to put those pages on the node where the task is running.
 315If the per-cpuset boolean flag file 'memory_spread_slab' is set,
 316then the kernel will spread some file system related slab caches,
 317such as for inodes and dentries evenly over all the nodes that the
 318faulting task is allowed to use, instead of preferring to put those
 319pages on the node where the task is running.
 321The setting of these flags does not affect anonymous data segment or
 322stack segment pages of a task.
 324By default, both kinds of memory spreading are off, and memory
 325pages are allocated on the node local to where the task is running,
 326except perhaps as modified by the tasks NUMA mempolicy or cpuset
 327configuration, so long as sufficient free memory pages are available.
 329When new cpusets are created, they inherit the memory spread settings
 330of their parent.
 332Setting memory spreading causes allocations for the affected page
 333or slab caches to ignore the tasks NUMA mempolicy and be spread
 334instead.    Tasks using mbind() or set_mempolicy() calls to set NUMA
 335mempolicies will not notice any change in these calls as a result of
 336their containing tasks memory spread settings.  If memory spreading
 337is turned off, then the currently specified NUMA mempolicy once again
 338applies to memory page allocations.
 340Both 'memory_spread_page' and 'memory_spread_slab' are boolean flag
 341files.  By default they contain "0", meaning that the feature is off
 342for that cpuset.  If a "1" is written to that file, then that turns
 343the named feature on.
 345The implementation is simple.
 347Setting the flag 'memory_spread_page' turns on a per-process flag
 348PF_SPREAD_PAGE for each task that is in that cpuset or subsequently
 349joins that cpuset.  The page allocation calls for the page cache
 350is modified to perform an inline check for this PF_SPREAD_PAGE task
 351flag, and if set, a call to a new routine cpuset_mem_spread_node()
 352returns the node to prefer for the allocation.
 354Similarly, setting 'memory_spread_slab' turns on the flag
 355PF_SPREAD_SLAB, and appropriately marked slab caches will allocate
 356pages from the node returned by cpuset_mem_spread_node().
 358The cpuset_mem_spread_node() routine is also simple.  It uses the
 359value of a per-task rotor cpuset_mem_spread_rotor to select the next
 360node in the current tasks mems_allowed to prefer for the allocation.
 362This memory placement policy is also known (in other contexts) as
 363round-robin or interleave.
 365This policy can provide substantial improvements for jobs that need
 366to place thread local data on the corresponding node, but that need
 367to access large file system data sets that need to be spread across
 368the several nodes in the jobs cpuset in order to fit.  Without this
 369policy, especially for jobs that might have one thread reading in the
 370data set, the memory allocation across the nodes in the jobs cpuset
 371can become very uneven.
 3731.7 What is sched_load_balance ?
 376The kernel scheduler (kernel/sched.c) automatically load balances
 377tasks.  If one CPU is underutilized, kernel code running on that
 378CPU will look for tasks on other more overloaded CPUs and move those
 379tasks to itself, within the constraints of such placement mechanisms
 380as cpusets and sched_setaffinity.
 382The algorithmic cost of load balancing and its impact on key shared
 383kernel data structures such as the task list increases more than
 384linearly with the number of CPUs being balanced.  So the scheduler
 385has support to partition the systems CPUs into a number of sched
 386domains such that it only load balances within each sched domain.
 387Each sched domain covers some subset of the CPUs in the system;
 388no two sched domains overlap; some CPUs might not be in any sched
 389domain and hence won't be load balanced.
 391Put simply, it costs less to balance between two smaller sched domains
 392than one big one, but doing so means that overloads in one of the
 393two domains won't be load balanced to the other one.
 395By default, there is one sched domain covering all CPUs, except those
 396marked isolated using the kernel boot time "isolcpus=" argument.
 398This default load balancing across all CPUs is not well suited for
 399the following two situations:
 400 1) On large systems, load balancing across many CPUs is expensive.
 401    If the system is managed using cpusets to place independent jobs
 402    on separate sets of CPUs, full load balancing is unnecessary.
 403 2) Systems supporting realtime on some CPUs need to minimize
 404    system overhead on those CPUs, including avoiding task load
 405    balancing if that is not needed.
 407When the per-cpuset flag "sched_load_balance" is enabled (the default
 408setting), it requests that all the CPUs in that cpusets allowed 'cpus'
 409be contained in a single sched domain, ensuring that load balancing
 410can move a task (not otherwised pinned, as by sched_setaffinity)
 411from any CPU in that cpuset to any other.
 413When the per-cpuset flag "sched_load_balance" is disabled, then the
 414scheduler will avoid load balancing across the CPUs in that cpuset,
 415--except-- in so far as is necessary because some overlapping cpuset
 416has "sched_load_balance" enabled.
 418So, for example, if the top cpuset has the flag "sched_load_balance"
 419enabled, then the scheduler will have one sched domain covering all
 420CPUs, and the setting of the "sched_load_balance" flag in any other
 421cpusets won't matter, as we're already fully load balancing.
 423Therefore in the above two situations, the top cpuset flag
 424"sched_load_balance" should be disabled, and only some of the smaller,
 425child cpusets have this flag enabled.
 427When doing this, you don't usually want to leave any unpinned tasks in
 428the top cpuset that might use non-trivial amounts of CPU, as such tasks
 429may be artificially constrained to some subset of CPUs, depending on
 430the particulars of this flag setting in descendent cpusets.  Even if
 431such a task could use spare CPU cycles in some other CPUs, the kernel
 432scheduler might not consider the possibility of load balancing that
 433task to that underused CPU.
 435Of course, tasks pinned to a particular CPU can be left in a cpuset
 436that disables "sched_load_balance" as those tasks aren't going anywhere
 437else anyway.
 439There is an impedance mismatch here, between cpusets and sched domains.
 440Cpusets are hierarchical and nest.  Sched domains are flat; they don't
 441overlap and each CPU is in at most one sched domain.
 443It is necessary for sched domains to be flat because load balancing
 444across partially overlapping sets of CPUs would risk unstable dynamics
 445that would be beyond our understanding.  So if each of two partially
 446overlapping cpusets enables the flag 'sched_load_balance', then we
 447form a single sched domain that is a superset of both.  We won't move
 448a task to a CPU outside it cpuset, but the scheduler load balancing
 449code might waste some compute cycles considering that possibility.
 451This mismatch is why there is not a simple one-to-one relation
 452between which cpusets have the flag "sched_load_balance" enabled,
 453and the sched domain configuration.  If a cpuset enables the flag, it
 454will get balancing across all its CPUs, but if it disables the flag,
 455it will only be assured of no load balancing if no other overlapping
 456cpuset enables the flag.
 458If two cpusets have partially overlapping 'cpus' allowed, and only
 459one of them has this flag enabled, then the other may find its
 460tasks only partially load balanced, just on the overlapping CPUs.
 461This is just the general case of the top_cpuset example given a few
 462paragraphs above.  In the general case, as in the top cpuset case,
 463don't leave tasks that might use non-trivial amounts of CPU in
 464such partially load balanced cpusets, as they may be artificially
 465constrained to some subset of the CPUs allowed to them, for lack of
 466load balancing to the other CPUs.
 4681.7.1 sched_load_balance implementation details.
 471The per-cpuset flag 'sched_load_balance' defaults to enabled (contrary
 472to most cpuset flags.)  When enabled for a cpuset, the kernel will
 473ensure that it can load balance across all the CPUs in that cpuset
 474(makes sure that all the CPUs in the cpus_allowed of that cpuset are
 475in the same sched domain.)
 477If two overlapping cpusets both have 'sched_load_balance' enabled,
 478then they will be (must be) both in the same sched domain.
 480If, as is the default, the top cpuset has 'sched_load_balance' enabled,
 481then by the above that means there is a single sched domain covering
 482the whole system, regardless of any other cpuset settings.
 484The kernel commits to user space that it will avoid load balancing
 485where it can.  It will pick as fine a granularity partition of sched
 486domains as it can while still providing load balancing for any set
 487of CPUs allowed to a cpuset having 'sched_load_balance' enabled.
 489The internal kernel cpuset to scheduler interface passes from the
 490cpuset code to the scheduler code a partition of the load balanced
 491CPUs in the system. This partition is a set of subsets (represented
 492as an array of struct cpumask) of CPUs, pairwise disjoint, that cover
 493all the CPUs that must be load balanced.
 495The cpuset code builds a new such partition and passes it to the
 496scheduler sched domain setup code, to have the sched domains rebuilt
 497as necessary, whenever:
 498 - the 'sched_load_balance' flag of a cpuset with non-empty CPUs changes,
 499 - or CPUs come or go from a cpuset with this flag enabled,
 500 - or 'sched_relax_domain_level' value of a cpuset with non-empty CPUs
 501   and with this flag enabled changes,
 502 - or a cpuset with non-empty CPUs and with this flag enabled is removed,
 503 - or a cpu is offlined/onlined.
 505This partition exactly defines what sched domains the scheduler should
 506setup - one sched domain for each element (struct cpumask) in the
 509The scheduler remembers the currently active sched domain partitions.
 510When the scheduler routine partition_sched_domains() is invoked from
 511the cpuset code to update these sched domains, it compares the new
 512partition requested with the current, and updates its sched domains,
 513removing the old and adding the new, for each change.
 5161.8 What is sched_relax_domain_level ?
 519In sched domain, the scheduler migrates tasks in 2 ways; periodic load
 520balance on tick, and at time of some schedule events.
 522When a task is woken up, scheduler try to move the task on idle CPU.
 523For example, if a task A running on CPU X activates another task B
 524on the same CPU X, and if CPU Y is X's sibling and performing idle,
 525then scheduler migrate task B to CPU Y so that task B can start on
 526CPU Y without waiting task A on CPU X.
 528And if a CPU run out of tasks in its runqueue, the CPU try to pull
 529extra tasks from other busy CPUs to help them before it is going to
 530be idle.
 532Of course it takes some searching cost to find movable tasks and/or
 533idle CPUs, the scheduler might not search all CPUs in the domain
 534everytime.  In fact, in some architectures, the searching ranges on
 535events are limited in the same socket or node where the CPU locates,
 536while the load balance on tick searchs all.
 538For example, assume CPU Z is relatively far from CPU X.  Even if CPU Z
 539is idle while CPU X and the siblings are busy, scheduler can't migrate
 540woken task B from X to Z since it is out of its searching range.
 541As the result, task B on CPU X need to wait task A or wait load balance
 542on the next tick.  For some applications in special situation, waiting
 5431 tick may be too long.
 545The 'sched_relax_domain_level' file allows you to request changing
 546this searching range as you like.  This file takes int value which
 547indicates size of searching range in levels ideally as follows,
 548otherwise initial value -1 that indicates the cpuset has no request.
 550  -1  : no request. use system default or follow request of others.
 551   0  : no search.
 552   1  : search siblings (hyperthreads in a core).
 553   2  : search cores in a package.
 554   3  : search cpus in a node [= system wide on non-NUMA system]
 555 ( 4  : search nodes in a chunk of node [on NUMA system] )
 556 ( 5  : search system wide [on NUMA system] )
 558The system default is architecture dependent.  The system default
 559can be changed using the relax_domain_level= boot parameter.
 561This file is per-cpuset and affect the sched domain where the cpuset
 562belongs to.  Therefore if the flag 'sched_load_balance' of a cpuset
 563is disabled, then 'sched_relax_domain_level' have no effect since
 564there is no sched domain belonging the cpuset.
 566If multiple cpusets are overlapping and hence they form a single sched
 567domain, the largest value among those is used.  Be careful, if one
 568requests 0 and others are -1 then 0 is used.
 570Note that modifying this file will have both good and bad effects,
 571and whether it is acceptable or not depends on your situation.
 572Don't modify this file if you are not sure.
 574If your situation is:
 575 - The migration costs between each cpu can be assumed considerably
 576   small(for you) due to your special application's behavior or
 577   special hardware support for CPU cache etc.
 578 - The searching cost doesn't have impact(for you) or you can make
 579   the searching cost enough small by managing cpuset to compact etc.
 580 - The latency is required even it sacrifices cache hit rate etc.
 581then increasing 'sched_relax_domain_level' would benefit you.
 5841.9 How do I use cpusets ?
 587In order to minimize the impact of cpusets on critical kernel
 588code, such as the scheduler, and due to the fact that the kernel
 589does not support one task updating the memory placement of another
 590task directly, the impact on a task of changing its cpuset CPU
 591or Memory Node placement, or of changing to which cpuset a task
 592is attached, is subtle.
 594If a cpuset has its Memory Nodes modified, then for each task attached
 595to that cpuset, the next time that the kernel attempts to allocate
 596a page of memory for that task, the kernel will notice the change
 597in the tasks cpuset, and update its per-task memory placement to
 598remain within the new cpusets memory placement.  If the task was using
 599mempolicy MPOL_BIND, and the nodes to which it was bound overlap with
 600its new cpuset, then the task will continue to use whatever subset
 601of MPOL_BIND nodes are still allowed in the new cpuset.  If the task
 602was using MPOL_BIND and now none of its MPOL_BIND nodes are allowed
 603in the new cpuset, then the task will be essentially treated as if it
 604was MPOL_BIND bound to the new cpuset (even though its numa placement,
 605as queried by get_mempolicy(), doesn't change).  If a task is moved
 606from one cpuset to another, then the kernel will adjust the tasks
 607memory placement, as above, the next time that the kernel attempts
 608to allocate a page of memory for that task.
 610If a cpuset has its 'cpus' modified, then each task in that cpuset
 611will have its allowed CPU placement changed immediately.  Similarly,
 612if a tasks pid is written to another cpusets 'tasks' file, then its
 613allowed CPU placement is changed immediately.  If such a task had been
 614bound to some subset of its cpuset using the sched_setaffinity() call,
 615the task will be allowed to run on any CPU allowed in its new cpuset,
 616negating the effect of the prior sched_setaffinity() call.
 618In summary, the memory placement of a task whose cpuset is changed is
 619updated by the kernel, on the next allocation of a page for that task,
 620and the processor placement is updated immediately.
 622Normally, once a page is allocated (given a physical page
 623of main memory) then that page stays on whatever node it
 624was allocated, so long as it remains allocated, even if the
 625cpusets memory placement policy 'mems' subsequently changes.
 626If the cpuset flag file 'memory_migrate' is set true, then when
 627tasks are attached to that cpuset, any pages that task had
 628allocated to it on nodes in its previous cpuset are migrated
 629to the tasks new cpuset. The relative placement of the page within
 630the cpuset is preserved during these migration operations if possible.
 631For example if the page was on the second valid node of the prior cpuset
 632then the page will be placed on the second valid node of the new cpuset.
 634Also if 'memory_migrate' is set true, then if that cpusets
 635'mems' file is modified, pages allocated to tasks in that
 636cpuset, that were on nodes in the previous setting of 'mems',
 637will be moved to nodes in the new setting of 'mems.'
 638Pages that were not in the tasks prior cpuset, or in the cpusets
 639prior 'mems' setting, will not be moved.
 641There is an exception to the above.  If hotplug functionality is used
 642to remove all the CPUs that are currently assigned to a cpuset,
 643then all the tasks in that cpuset will be moved to the nearest ancestor
 644with non-empty cpus.  But the moving of some (or all) tasks might fail if
 645cpuset is bound with another cgroup subsystem which has some restrictions
 646on task attaching.  In this failing case, those tasks will stay
 647in the original cpuset, and the kernel will automatically update
 648their cpus_allowed to allow all online CPUs.  When memory hotplug
 649functionality for removing Memory Nodes is available, a similar exception
 650is expected to apply there as well.  In general, the kernel prefers to
 651violate cpuset placement, over starving a task that has had all
 652its allowed CPUs or Memory Nodes taken offline.
 654There is a second exception to the above.  GFP_ATOMIC requests are
 655kernel internal allocations that must be satisfied, immediately.
 656The kernel may drop some request, in rare cases even panic, if a
 657GFP_ATOMIC alloc fails.  If the request cannot be satisfied within
 658the current tasks cpuset, then we relax the cpuset, and look for
 659memory anywhere we can find it.  It's better to violate the cpuset
 660than stress the kernel.
 662To start a new job that is to be contained within a cpuset, the steps are:
 664 1) mkdir /dev/cpuset
 665 2) mount -t cgroup -ocpuset cpuset /dev/cpuset
 666 3) Create the new cpuset by doing mkdir's and write's (or echo's) in
 667    the /dev/cpuset virtual file system.
 668 4) Start a task that will be the "founding father" of the new job.
 669 5) Attach that task to the new cpuset by writing its pid to the
 670    /dev/cpuset tasks file for that cpuset.
 671 6) fork, exec or clone the job tasks from this founding father task.
 673For example, the following sequence of commands will setup a cpuset
 674named "Charlie", containing just CPUs 2 and 3, and Memory Node 1,
 675and then start a subshell 'sh' in that cpuset:
 677  mount -t cgroup -ocpuset cpuset /dev/cpuset
 678  cd /dev/cpuset
 679  mkdir Charlie
 680  cd Charlie
 681  /bin/echo 2-3 > cpus
 682  /bin/echo 1 > mems
 683  /bin/echo $$ > tasks
 684  sh
 685  # The subshell 'sh' is now running in cpuset Charlie
 686  # The next line should display '/Charlie'
 687  cat /proc/self/cpuset
 689There are ways to query or modify cpusets:
 690 - via the cpuset file system directly, using the various cd, mkdir, echo,
 691   cat, rmdir commands from the shell, or their equivalent from C.
 692 - via the C library libcpuset.
 693 - via the C library libcgroup.
 694   (
 695 - via the python application cset.
 696   (
 698The sched_setaffinity calls can also be done at the shell prompt using
 699SGI's runon or Robert Love's taskset.  The mbind and set_mempolicy
 700calls can be done at the shell prompt using the numactl command
 701(part of Andi Kleen's numa package).
 7032. Usage Examples and Syntax
 7062.1 Basic Usage
 709Creating, modifying, using the cpusets can be done through the cpuset
 710virtual filesystem.
 712To mount it, type:
 713# mount -t cgroup -o cpuset cpuset /dev/cpuset
 715Then under /dev/cpuset you can find a tree that corresponds to the
 716tree of the cpusets in the system. For instance, /dev/cpuset
 717is the cpuset that holds the whole system.
 719If you want to create a new cpuset under /dev/cpuset:
 720# cd /dev/cpuset
 721# mkdir my_cpuset
 723Now you want to do something with this cpuset.
 724# cd my_cpuset
 726In this directory you can find several files:
 727# ls
 728cpu_exclusive  memory_migrate      mems                      tasks
 729cpus           memory_pressure     notify_on_release
 730mem_exclusive  memory_spread_page  sched_load_balance
 731mem_hardwall   memory_spread_slab  sched_relax_domain_level
 733Reading them will give you information about the state of this cpuset:
 734the CPUs and Memory Nodes it can use, the processes that are using
 735it, its properties.  By writing to these files you can manipulate
 736the cpuset.
 738Set some flags:
 739# /bin/echo 1 > cpu_exclusive
 741Add some cpus:
 742# /bin/echo 0-7 > cpus
 744Add some mems:
 745# /bin/echo 0-7 > mems
 747Now attach your shell to this cpuset:
 748# /bin/echo $$ > tasks
 750You can also create cpusets inside your cpuset by using mkdir in this
 752# mkdir my_sub_cs
 754To remove a cpuset, just use rmdir:
 755# rmdir my_sub_cs
 756This will fail if the cpuset is in use (has cpusets inside, or has
 757processes attached).
 759Note that for legacy reasons, the "cpuset" filesystem exists as a
 760wrapper around the cgroup filesystem.
 762The command
 764mount -t cpuset X /dev/cpuset
 766is equivalent to
 768mount -t cgroup -ocpuset,noprefix X /dev/cpuset
 769echo "/sbin/cpuset_release_agent" > /dev/cpuset/release_agent
 7712.2 Adding/removing cpus
 774This is the syntax to use when writing in the cpus or mems files
 775in cpuset directories:
 777# /bin/echo 1-4 > cpus          -> set cpus list to cpus 1,2,3,4
 778# /bin/echo 1,2,3,4 > cpus      -> set cpus list to cpus 1,2,3,4
 7802.3 Setting flags
 783The syntax is very simple:
 785# /bin/echo 1 > cpu_exclusive   -> set flag 'cpu_exclusive'
 786# /bin/echo 0 > cpu_exclusive   -> unset flag 'cpu_exclusive'
 7882.4 Attaching processes
 791# /bin/echo PID > tasks
 793Note that it is PID, not PIDs. You can only attach ONE task at a time.
 794If you have several tasks to attach, you have to do it one after another:
 796# /bin/echo PID1 > tasks
 797# /bin/echo PID2 > tasks
 798        ...
 799# /bin/echo PIDn > tasks
 8023. Questions
 805Q: what's up with this '/bin/echo' ?
 806A: bash's builtin 'echo' command does not check calls to write() against
 807   errors. If you use it in the cpuset file system, you won't be
 808   able to tell whether a command succeeded or failed.
 810Q: When I attach processes, only the first of the line gets really attached !
 811A: We can only return one error code per call to write(). So you should also
 812   put only ONE pid.
 8144. Contact