In addition to controlling which cpus and mems a process is
allowed to use, cpusets provide the following extended
capabilities.
Exclusive cpusets
If a cpuset is marked cpu_exclusive or mem_exclusive, no other
cpuset, other than a direct ancestor or descendant, may share any
of the same CPUs or memory nodes.
A cpuset that is mem_exclusive restricts kernel allocations for
buffer cache pages and other internal kernel data pages commonly
shared by the kernel across multiple users. All cpusets, whether
mem_exclusive or not, restrict allocations of memory for user
space. This enables configuring a system so that several
independent jobs can share common kernel data, while isolating
each job's user allocation in its own cpuset. To do this,
construct a large mem_exclusive cpuset to hold all the jobs, and
construct child, non-mem_exclusive cpusets for each individual
job. Only a small amount of kernel memory, such as requests from
interrupt handlers, is allowed to be placed on memory nodes
outside even a mem_exclusive cpuset.
Hardwall
A cpuset that has mem_exclusive or mem_hardwall set is a hardwall
cpuset. A hardwall cpuset restricts kernel allocations for page,
buffer, and other data commonly shared by the kernel across
multiple users. All cpusets, whether hardwall or not, restrict
allocations of memory for user space.
This enables configuring a system so that several independent
jobs can share common kernel data, such as filesystem pages,
while isolating each job's user allocation in its own cpuset. To
do this, construct a large hardwall cpuset to hold all the jobs,
and construct child cpusets for each individual job which are not
hardwall cpusets.
Only a small amount of kernel memory, such as requests from
interrupt handlers, is allowed to be taken outside even a
hardwall cpuset.
Notify on release
If the notify_on_release flag is enabled (1) in a cpuset, then
whenever the last process in the cpuset leaves (exits or attaches
to some other cpuset) and the last child cpuset of that cpuset is
removed, the kernel will run the command
/sbin/cpuset_release_agent, supplying the pathname (relative to
the mount point of the cpuset filesystem) of the abandoned
cpuset. This enables automatic removal of abandoned cpusets.
The default value of notify_on_release in the root cpuset at
system boot is disabled (0). The default value of other cpusets
at creation is the current value of their parent's
notify_on_release setting.
The command /sbin/cpuset_release_agent is invoked, with the name
(/dev/cpuset relative path) of the to-be-released cpuset in
argv[1].
The usual contents of the command /sbin/cpuset_release_agent is
simply the shell script:
#!/bin/sh
rmdir /dev/cpuset/$1
As with other flag values below, this flag can be changed by
writing an ASCII number 0 or 1 (with optional trailing newline)
into the file, to clear or set the flag, respectively.
Memory pressure
The memory_pressure of a cpuset provides a simple per-cpuset
running average of the rate that the processes in a cpuset are
attempting to free up in-use memory on the nodes of the cpuset to
satisfy additional memory requests.
This enables batch managers that are monitoring jobs running in
dedicated cpusets to efficiently detect what level of memory
pressure that job is causing.
This is useful both on tightly managed systems running a wide mix
of submitted jobs, which may choose to terminate or reprioritize
jobs that are trying to use more memory than allowed on the nodes
assigned them, and with tightly coupled, long-running, massively
parallel scientific computing jobs that will dramatically fail to
meet required performance goals if they start to use more memory
than allowed to them.
This mechanism provides a very economical way for the batch
manager to monitor a cpuset for signs of memory pressure. It's
up to the batch manager or other user code to decide what action
to take if it detects signs of memory pressure.
Unless memory pressure calculation is enabled by setting the
pseudo-file /dev/cpuset/cpuset.memory_pressure_enabled, it is not
computed for any cpuset, and reads from any memory_pressure
always return zero, as represented by the ASCII string "0\n".
See the WARNINGS section, below.
A per-cpuset, running average is employed for the following
reasons:
* Because this meter is per-cpuset rather than per-process or
per virtual memory region, the system load imposed by a batch
scheduler monitoring this metric is sharply reduced on large
systems, because a scan of the tasklist can be avoided on each
set of queries.
* Because this meter is a running average rather than an
accumulating counter, a batch scheduler can detect memory
pressure with a single read, instead of having to read and
accumulate results for a period of time.
* Because this meter is per-cpuset rather than per-process, the
batch scheduler can obtain the key information—memory pressure
in a cpuset—with a single read, rather than having to query
and accumulate results over all the (dynamically changing) set
of processes in the cpuset.
The memory_pressure of a cpuset is calculated using a per-cpuset
simple digital filter that is kept within the kernel. For each
cpuset, this filter tracks the recent rate at which processes
attached to that cpuset enter the kernel direct reclaim code.
The kernel direct reclaim code is entered whenever a process has
to satisfy a memory page request by first finding some other page
to repurpose, due to lack of any readily available already free
pages. Dirty filesystem pages are repurposed by first writing
them to disk. Unmodified filesystem buffer pages are repurposed
by simply dropping them, though if that page is needed again, it
will have to be reread from disk.
The cpuset.memory_pressure file provides an integer number
representing the recent (half-life of 10 seconds) rate of entries
to the direct reclaim code caused by any process in the cpuset,
in units of reclaims attempted per second, times 1000.
Memory spread
There are two Boolean flag files per cpuset that control where
the kernel allocates pages for the filesystem buffers and related
in-kernel data structures. They are called
cpuset.memory_spread_page and cpuset.memory_spread_slab.
If the per-cpuset Boolean flag file cpuset.memory_spread_page is
set, then the kernel will spread the filesystem buffers (page
cache) evenly over all the nodes that the faulting process is
allowed to use, instead of preferring to put those pages on the
node where the process is running.
If the per-cpuset Boolean flag file cpuset.memory_spread_slab is
set, then the kernel will spread some filesystem-related slab
caches, such as those for inodes and directory entries, evenly
over all the nodes that the faulting process is allowed to use,
instead of preferring to put those pages on the node where the
process is running.
The setting of these flags does not affect the data segment (see
brk(2)) or stack segment pages of a process.
By default, both kinds of memory spreading are off and the kernel
prefers to allocate memory pages on the node local to where the
requesting process is running. If that node is not allowed by
the process's NUMA memory policy or cpuset configuration or if
there are insufficient free memory pages on that node, then the
kernel looks for the nearest node that is allowed and has
sufficient free memory.
When new cpusets are created, they inherit the memory spread
settings of their parent.
Setting memory spreading causes allocations for the affected page
or slab caches to ignore the process's NUMA memory policy and be
spread instead. However, the effect of these changes in memory
placement caused by cpuset-specified memory spreading is hidden
from the mbind(2) or set_mempolicy(2) calls. These two NUMA
memory policy calls always appear to behave as if no cpuset-
specified memory spreading is in effect, even if it is. If
cpuset memory spreading is subsequently turned off, the NUMA
memory policy most recently specified by these calls is
automatically reapplied.
Both cpuset.memory_spread_page and cpuset.memory_spread_slab are
Boolean flag files. By default, they contain "0", meaning that
the feature is off for that cpuset. If a "1" is written to that
file, that turns the named feature on.
Cpuset-specified memory spreading behaves similarly to what is
known (in other contexts) as round-robin or interleave memory
placement.
Cpuset-specified memory spreading can provide substantial
performance improvements for jobs that:
a) need to place thread-local data on memory nodes close to the
CPUs which are running the threads that most frequently access
that data; but also
b) need to access large filesystem data sets that must to be
spread across the several nodes in the job's cpuset in order
to fit.
Without this policy, the memory allocation across the nodes in
the job's cpuset can become very uneven, especially for jobs that
might have just a single thread initializing or reading in the
data set.
Memory migration
Normally, under the default setting (disabled) of
cpuset.memory_migrate, once a page is allocated (given a physical
page of main memory), then that page stays on whatever node it
was allocated, so long as it remains allocated, even if the
cpuset's memory-placement policy mems subsequently changes.
When memory migration is enabled in a cpuset, if the mems setting
of the cpuset is changed, then any memory page in use by any
process in the cpuset that is on a memory node that is no longer
allowed will be migrated to a memory node that is allowed.
Furthermore, if a process is moved into a cpuset with
memory_migrate enabled, any memory pages it uses that were on
memory nodes allowed in its previous cpuset, but which are not
allowed in its new cpuset, will be migrated to a memory node
allowed in the new cpuset.
The relative placement of a migrated page within the cpuset is
preserved during these migration operations if possible. For
example, if the page was on the second valid node of the prior
cpuset, then the page will be placed on the second valid node of
the new cpuset, if possible.
Scheduler load balancing
The kernel scheduler automatically load balances processes. If
one CPU is underutilized, the kernel will look for processes on
other more overloaded CPUs and move those processes to the
underutilized CPU, within the constraints of such placement
mechanisms as cpusets and sched_setaffinity(2).
The algorithmic cost of load balancing and its impact on key
shared kernel data structures such as the process list increases
more than linearly with the number of CPUs being balanced. For
example, it costs more to load balance across one large set of
CPUs than it does to balance across two smaller sets of CPUs,
each of half the size of the larger set. (The precise
relationship between the number of CPUs being balanced and the
cost of load balancing depends on implementation details of the
kernel process scheduler, which is subject to change over time,
as improved kernel scheduler algorithms are implemented.)
The per-cpuset flag sched_load_balance provides a mechanism to
suppress this automatic scheduler load balancing in cases where
it is not needed and suppressing it would have worthwhile
performance benefits.
By default, load balancing is done across all CPUs, except those
marked isolated using the kernel boot time "isolcpus=" argument.
(See Scheduler Relax Domain Level, below, to change this
default.)
This default load balancing across all CPUs is not well suited to
the following two situations:
* On large systems, load balancing across many CPUs is
expensive. If the system is managed using cpusets to place
independent jobs on separate sets of CPUs, full load balancing
is unnecessary.
* Systems supporting real-time on some CPUs need to minimize
system overhead on those CPUs, including avoiding process load
balancing if that is not needed.
When the per-cpuset flag sched_load_balance is enabled (the
default setting), it requests load balancing across all the CPUs
in that cpuset's allowed CPUs, ensuring that load balancing can
move a process (not otherwise pinned, as by sched_setaffinity(2))
from any CPU in that cpuset to any other.
When the per-cpuset flag sched_load_balance is disabled, then the
scheduler will avoid load balancing across the CPUs in that
cpuset, except in so far as is necessary because some overlapping
cpuset has sched_load_balance enabled.
So, for example, if the top cpuset has the flag
sched_load_balance enabled, then the scheduler will load balance
across all CPUs, and the setting of the sched_load_balance flag
in other cpusets has no effect, as we're already fully load
balancing.
Therefore in the above two situations, the flag
sched_load_balance should be disabled in the top cpuset, and only
some of the smaller, child cpusets would have this flag enabled.
When doing this, you don't usually want to leave any unpinned
processes in the top cpuset that might use nontrivial amounts of
CPU, as such processes may be artificially constrained to some
subset of CPUs, depending on the particulars of this flag setting
in descendant cpusets. Even if such a process could use spare
CPU cycles in some other CPUs, the kernel scheduler might not
consider the possibility of load balancing that process to the
underused CPU.
Of course, processes pinned to a particular CPU can be left in a
cpuset that disables sched_load_balance as those processes aren't
going anywhere else anyway.
Scheduler relax domain level
The kernel scheduler performs immediate load balancing whenever a
CPU becomes free or another task becomes runnable. This load
balancing works to ensure that as many CPUs as possible are
usefully employed running tasks. The kernel also performs
periodic load balancing off the software clock described in
time(7). The setting of sched_relax_domain_level applies only to
immediate load balancing. Regardless of the
sched_relax_domain_level setting, periodic load balancing is
attempted over all CPUs (unless disabled by turning off
sched_load_balance.) In any case, of course, tasks will be
scheduled to run only on CPUs allowed by their cpuset, as
modified by sched_setaffinity(2) system calls.
On small systems, such as those with just a few CPUs, immediate
load balancing is useful to improve system interactivity and to
minimize wasteful idle CPU cycles. But on large systems,
attempting immediate load balancing across a large number of CPUs
can be more costly than it is worth, depending on the particular
performance characteristics of the job mix and the hardware.
The exact meaning of the small integer values of
sched_relax_domain_level will depend on internal implementation
details of the kernel scheduler code and on the non-uniform
architecture of the hardware. Both of these will evolve over
time and vary by system architecture and kernel version.
As of this writing, when this capability was introduced in Linux
2.6.26, on certain popular architectures, the positive values of
sched_relax_domain_level have the following meanings.
(1) Perform immediate load balancing across Hyper-Thread siblings
on the same core.
(2) Perform immediate load balancing across other cores in the
same package.
(3) Perform immediate load balancing across other CPUs on the
same node or blade.
(4) Perform immediate load balancing across over several
(implementation detail) nodes [On NUMA systems].
(5) Perform immediate load balancing across over all CPUs in
system [On NUMA systems].
The sched_relax_domain_level value of zero (0) always means don't
perform immediate load balancing, hence that load balancing is
done only periodically, not immediately when a CPU becomes
available or another task becomes runnable.
The sched_relax_domain_level value of minus one (-1) always means
use the system default value. The system default value can vary
by architecture and kernel version. This system default value
can be changed by kernel boot-time "relax_domain_level="
argument.
In the case of multiple overlapping cpusets which have
conflicting sched_relax_domain_level values, then the highest
such value applies to all CPUs in any of the overlapping cpusets.
In such cases, the value minus one (-1) is the lowest value,
overridden by any other value, and the value zero (0) is the next
lowest value.
