CPU binding

Enumerations
enum	hwloc_cpubind_flags_t { HWLOC_CPUBIND_PROCESS , HWLOC_CPUBIND_THREAD , HWLOC_CPUBIND_STRICT , HWLOC_CPUBIND_NOMEMBIND }

Functions
int	hwloc_set_cpubind (hwloc_topology_t topology, hwloc_const_cpuset_t set, int flags)
int	hwloc_get_cpubind (hwloc_topology_t topology, hwloc_cpuset_t set, int flags)
int	hwloc_set_proc_cpubind (hwloc_topology_t topology, hwloc_pid_t pid, hwloc_const_cpuset_t set, int flags)
int	hwloc_get_proc_cpubind (hwloc_topology_t topology, hwloc_pid_t pid, hwloc_cpuset_t set, int flags)
int	hwloc_set_thread_cpubind (hwloc_topology_t topology, hwloc_thread_t thread, hwloc_const_cpuset_t set, int flags)
int	hwloc_get_thread_cpubind (hwloc_topology_t topology, hwloc_thread_t thread, hwloc_cpuset_t set, int flags)
int	hwloc_get_last_cpu_location (hwloc_topology_t topology, hwloc_cpuset_t set, int flags)
int	hwloc_get_proc_last_cpu_location (hwloc_topology_t topology, hwloc_pid_t pid, hwloc_cpuset_t set, int flags)

Detailed Description

Some operating systems only support binding threads or processes to a single PU. Others allow binding to larger sets such as entire Cores or Packages or even random sets of individual PUs. In such operating system, the scheduler is free to run the task on one of these PU, then migrate it to another PU, etc. It is often useful to call hwloc_bitmap_singlify() on the target CPU set before passing it to the binding function to avoid these expensive migrations. See the documentation of hwloc_bitmap_singlify() for details.

Some operating systems do not provide all hwloc-supported mechanisms to bind processes, threads, etc. hwloc_topology_get_support() may be used to query about the actual CPU binding support in the currently used operating system.

When the requested binding operation is not available and the HWLOC_CPUBIND_STRICT flag was passed, the function returns -1. errno is set to ENOSYS when it is not possible to bind the requested kind of object processes/threads. errno is set to EXDEV when the requested cpuset can not be enforced (e.g. some systems only allow one CPU, and some other systems only allow one NUMA node).

If HWLOC_CPUBIND_STRICT was not passed, the function may fail as well, or the operating system may use a slightly different operation (with side-effects, smaller binding set, etc.) when the requested operation is not exactly supported.

The most portable version that should be preferred over the others, whenever possible, is the following one which just binds the current program, assuming it is single-threaded:

hwloc_set_cpubind(topology, set, 0),

If the program may be multithreaded, the following one should be preferred to only bind the current thread:

hwloc_set_cpubind(topology, set, HWLOC_CPUBIND_THREAD),

See also

Some example codes are available under doc/examples/ in the source tree.

Note

To unbind, just call the binding function with either a full cpuset or a cpuset equal to the system cpuset.

On some operating systems, CPU binding may have effects on memory binding, see HWLOC_CPUBIND_NOMEMBIND

Running lstopo --top or hwloc-ps can be a very convenient tool to check how binding actually happened.

Enumeration Type Documentation

hwloc_cpubind_flags_t

enum hwloc_cpubind_flags_t

Process/Thread binding flags.

These bit flags can be used to refine the binding policy.

The default (0) is to bind the current process, assumed to be single-threaded, in a non-strict way. This is the most portable way to bind as all operating systems usually provide it.

Note

Not all systems support all kinds of binding. See the "Detailed Description" section of CPU binding for a description of errors that can occur.

Enumerator
HWLOC_CPUBIND_PROCESS	Bind all threads of the current (possibly) multithreaded process.
HWLOC_CPUBIND_THREAD	Bind current thread of current process.
HWLOC_CPUBIND_STRICT	Request for strict binding from the OS. By default, when the designated CPUs are all busy while other CPUs are idle, operating systems may execute the thread/process on those other CPUs instead of the designated CPUs, to let them progress anyway. Strict binding means that the thread/process will _never_ execute on other CPUs than the designated CPUs, even when those are busy with other tasks and other CPUs are idle. Note Depending on the operating system, strict binding may not be possible (e.g., the OS does not implement it) or not allowed (e.g., for an administrative reasons), and the function will fail in that case. When retrieving the binding of a process, this flag checks whether all its threads actually have the same binding. If the flag is not given, the binding of each thread will be accumulated. Note This flag is meaningless when retrieving the binding of a thread.
HWLOC_CPUBIND_NOMEMBIND	Avoid any effect on memory binding. On some operating systems, some CPU binding function would also bind the memory on the corresponding NUMA node. It is often not a problem for the application, but if it is, setting this flag will make hwloc avoid using OS functions that would also bind memory. This will however reduce the support of CPU bindings, i.e. potentially return -1 with errno set to ENOSYS in some cases. This flag is only meaningful when used with functions that set the CPU binding. It is ignored when used with functions that get CPU binding information.

Function Documentation

hwloc_get_cpubind()

int hwloc_get_cpubind	(	hwloc_topology_t	topology,
		hwloc_cpuset_t	set,
		int	flags
	)

Get current process or thread binding.

The CPU-set set (previously allocated by the caller) is filled with the list of PUs which the process or thread (according to flags) was last bound to.

Returns

0 on success, -1 on error.

hwloc_get_last_cpu_location()

int hwloc_get_last_cpu_location	(	hwloc_topology_t	topology,
		hwloc_cpuset_t	set,
		int	flags
	)

Get the last physical CPU where the current process or thread ran.

The CPU-set set (previously allocated by the caller) is filled with the list of PUs which the process or thread (according to flags) last ran on.

The operating system may move some tasks from one processor to another at any time according to their binding, so this function may return something that is already outdated.

flags can include either HWLOC_CPUBIND_PROCESS or HWLOC_CPUBIND_THREAD to specify whether the query should be for the whole process (union of all CPUs on which all threads are running), or only the current thread. If the process is single-threaded, flags can be set to zero to let hwloc use whichever method is available on the underlying OS.

Returns

0 on success, -1 on error.

hwloc_get_proc_cpubind()

int hwloc_get_proc_cpubind	(	hwloc_topology_t	topology,
		hwloc_pid_t	pid,
		hwloc_cpuset_t	set,
		int	flags
	)

Get the current physical binding of process pid.

The CPU-set set (previously allocated by the caller) is filled with the list of PUs which the process was last bound to.

Returns

0 on success, -1 on error.

Note

hwloc_pid_t is pid_t on Unix platforms, and HANDLE on native Windows platforms.

As a special case on Linux, if a tid (thread ID) is supplied instead of a pid (process ID) and HWLOC_CPUBIND_THREAD is passed in flags, the binding for that specific thread is returned.

On non-Linux systems, HWLOC_CPUBIND_THREAD can not be used in flags.

hwloc_get_proc_last_cpu_location()

int hwloc_get_proc_last_cpu_location	(	hwloc_topology_t	topology,
		hwloc_pid_t	pid,
		hwloc_cpuset_t	set,
		int	flags
	)

Get the last physical CPU where a process ran.

The CPU-set set (previously allocated by the caller) is filled with the list of PUs which the process last ran on.

The operating system may move some tasks from one processor to another at any time according to their binding, so this function may return something that is already outdated.

Returns

0 on success, -1 on error.

Note

hwloc_pid_t is pid_t on Unix platforms, and HANDLE on native Windows platforms.

As a special case on Linux, if a tid (thread ID) is supplied instead of a pid (process ID) and HWLOC_CPUBIND_THREAD is passed in flags, the last CPU location of that specific thread is returned.

On non-Linux systems, HWLOC_CPUBIND_THREAD can not be used in flags.

hwloc_get_thread_cpubind()

int hwloc_get_thread_cpubind	(	hwloc_topology_t	topology,
		hwloc_thread_t	thread,
		hwloc_cpuset_t	set,
		int	flags
	)

Get the current physical binding of thread tid.

The CPU-set set (previously allocated by the caller) is filled with the list of PUs which the thread was last bound to.

Returns

0 on success, -1 on error.

Note

hwloc_thread_t is pthread_t on Unix platforms, and HANDLE on native Windows platforms.

HWLOC_CPUBIND_PROCESS can not be used in flags.

hwloc_set_cpubind()

int hwloc_set_cpubind	(	hwloc_topology_t	topology,
		hwloc_const_cpuset_t	set,
		int	flags
	)

Bind current process or thread on CPUs given in physical bitmap set.

Returns

0 on success.

-1 with errno set to ENOSYS if the action is not supported.

-1 with errno set to EXDEV if the binding cannot be enforced.

hwloc_set_proc_cpubind()

int hwloc_set_proc_cpubind	(	hwloc_topology_t	topology,
		hwloc_pid_t	pid,
		hwloc_const_cpuset_t	set,
		int	flags
	)

Bind a process pid on CPUs given in physical bitmap set.

Returns

0 on success, -1 on error.

Note

hwloc_pid_t is pid_t on Unix platforms, and HANDLE on native Windows platforms.

As a special case on Linux, if a tid (thread ID) is supplied instead of a pid (process ID) and HWLOC_CPUBIND_THREAD is passed in flags, the binding is applied to that specific thread.

On non-Linux systems, HWLOC_CPUBIND_THREAD can not be used in flags.

hwloc_set_thread_cpubind()

int hwloc_set_thread_cpubind	(	hwloc_topology_t	topology,
		hwloc_thread_t	thread,
		hwloc_const_cpuset_t	set,
		int	flags
	)

Bind a thread thread on CPUs given in physical bitmap set.

Returns

0 on success, -1 on error.

Note

hwloc_thread_t is pthread_t on Unix platforms, and HANDLE on native Windows platforms.

HWLOC_CPUBIND_PROCESS can not be used in flags.