Memory binding

Enumerations
enum	hwloc_membind_policy_t {   HWLOC_MEMBIND_DEFAULT , HWLOC_MEMBIND_FIRSTTOUCH , HWLOC_MEMBIND_BIND , HWLOC_MEMBIND_INTERLEAVE ,   HWLOC_MEMBIND_NEXTTOUCH , HWLOC_MEMBIND_MIXED }
enum	hwloc_membind_flags_t {   HWLOC_MEMBIND_PROCESS , HWLOC_MEMBIND_THREAD , HWLOC_MEMBIND_STRICT , HWLOC_MEMBIND_MIGRATE ,   HWLOC_MEMBIND_NOCPUBIND , HWLOC_MEMBIND_BYNODESET }

Functions
int	hwloc_set_membind (hwloc_topology_t topology, hwloc_const_bitmap_t set, hwloc_membind_policy_t policy, int flags)
int	hwloc_get_membind (hwloc_topology_t topology, hwloc_bitmap_t set, hwloc_membind_policy_t *policy, int flags)
int	hwloc_set_proc_membind (hwloc_topology_t topology, hwloc_pid_t pid, hwloc_const_bitmap_t set, hwloc_membind_policy_t policy, int flags)
int	hwloc_get_proc_membind (hwloc_topology_t topology, hwloc_pid_t pid, hwloc_bitmap_t set, hwloc_membind_policy_t *policy, int flags)
int	hwloc_set_area_membind (hwloc_topology_t topology, const void *addr, size_t len, hwloc_const_bitmap_t set, hwloc_membind_policy_t policy, int flags)
int	hwloc_get_area_membind (hwloc_topology_t topology, const void *addr, size_t len, hwloc_bitmap_t set, hwloc_membind_policy_t *policy, int flags)
int	hwloc_get_area_memlocation (hwloc_topology_t topology, const void *addr, size_t len, hwloc_bitmap_t set, int flags)
void *	hwloc_alloc (hwloc_topology_t topology, size_t len)
void *	hwloc_alloc_membind (hwloc_topology_t topology, size_t len, hwloc_const_bitmap_t set, hwloc_membind_policy_t policy, int flags)
static void *	hwloc_alloc_membind_policy (hwloc_topology_t topology, size_t len, hwloc_const_bitmap_t set, hwloc_membind_policy_t policy, int flags)
int	hwloc_free (hwloc_topology_t topology, void *addr, size_t len)

Detailed Description

Memory binding can be done three ways:

• explicit memory allocation thanks to hwloc_alloc_membind() and friends: the binding will have effect on the memory allocated by these functions.
• implicit memory binding through binding policy: hwloc_set_membind() and friends only define the current policy of the process, which will be applied to the subsequent calls to malloc() and friends.
• migration of existing memory ranges, thanks to hwloc_set_area_membind() and friends, which move already-allocated data.

Not all operating systems support all three ways. hwloc_topology_get_support() may be used to query about the actual memory binding support in the currently used operating system.

When the requested binding operation is not available and the HWLOC_MEMBIND_STRICT flag was passed, the function returns -1. errno will be set to ENOSYS when the system does support the specified action or policy (e.g., some systems only allow binding memory on a per-thread basis, whereas other systems only allow binding memory for all threads in a process). errno will be set to EXDEV when the requested set can not be enforced (e.g., some systems only allow binding memory to a single NUMA node).

If HWLOC_MEMBIND_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 form that should be preferred over the others whenever possible is as follows. It allocates some memory hopefully bound to the specified set. To do so, hwloc will possibly have to change the current memory binding policy in order to actually get the memory bound, if the OS does not provide any other way to simply allocate bound memory without changing the policy for all allocations. That is the difference with hwloc_alloc_membind(), which will never change the current memory binding policy.

hwloc_alloc_membind_policy(topology, size, set,
                           HWLOC_MEMBIND_BIND, 0);

Each hwloc memory binding function takes a bitmap argument that is a CPU set by default, or a NUMA memory node set if the flag HWLOC_MEMBIND_BYNODESET is specified. See Object Sets (hwloc_cpuset_t and hwloc_nodeset_t) and The bitmap API for a discussion of CPU sets and NUMA memory node sets. It is also possible to convert between CPU set and node set using hwloc_cpuset_to_nodeset() or hwloc_cpuset_from_nodeset().

Memory binding by CPU set cannot work for CPU-less NUMA memory nodes. Binding by nodeset should therefore be preferred whenever possible.

See also

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

Note

On some operating systems, memory binding affects the CPU binding; see HWLOC_MEMBIND_NOCPUBIND

Enumeration Type Documentation

hwloc_membind_flags_t

enum hwloc_membind_flags_t

Memory binding flags.

These flags can be used to refine the binding policy. All flags can be logically OR'ed together with the exception of HWLOC_MEMBIND_PROCESS and HWLOC_MEMBIND_THREAD; these two flags are mutually exclusive.

Not all systems support all kinds of binding. hwloc_topology_get_support() may be used to query about the actual memory binding support in the currently used operating system. See the "Detailed Description" section of Memory binding for a description of errors that can occur.

Enumerator
HWLOC_MEMBIND_PROCESS	Set policy for all threads of the specified (possibly multithreaded) process. This flag is mutually exclusive with HWLOC_MEMBIND_THREAD.
HWLOC_MEMBIND_THREAD	Set policy for a specific thread of the current process. This flag is mutually exclusive with HWLOC_MEMBIND_PROCESS.
HWLOC_MEMBIND_STRICT	Request strict binding from the OS. The function will fail if the binding can not be guaranteed / completely enforced. This flag has slightly different meanings depending on which function it is used with.
HWLOC_MEMBIND_MIGRATE	Migrate existing allocated memory. If the memory cannot be migrated and the HWLOC_MEMBIND_STRICT flag is passed, an error will be returned.
HWLOC_MEMBIND_NOCPUBIND	Avoid any effect on CPU binding. On some operating systems, some underlying memory binding functions also bind the application to the corresponding CPU(s). Using this flag will cause hwloc to avoid using OS functions that could potentially affect CPU bindings. Note, however, that using NOCPUBIND may reduce hwloc's overall memory binding support. Specifically: some of hwloc's memory binding functions may fail with errno set to ENOSYS when used with NOCPUBIND.
HWLOC_MEMBIND_BYNODESET	Consider the bitmap argument as a nodeset. The bitmap argument is considered a nodeset if this flag is given, or a cpuset otherwise by default. Memory binding by CPU set cannot work for CPU-less NUMA memory nodes. Binding by nodeset should therefore be preferred whenever possible.

hwloc_membind_policy_t

enum hwloc_membind_policy_t

Memory binding policy.

These constants can be used to choose the binding policy. Only one policy can be used at a time (i.e., the values cannot be OR'ed together).

Not all systems support all kinds of binding. hwloc_topology_get_support() may be used to query about the actual memory binding policy support in the currently used operating system. See the "Detailed Description" section of Memory binding for a description of errors that can occur.

Enumerator
HWLOC_MEMBIND_DEFAULT	Reset the memory allocation policy to the system default. Depending on the operating system, this may correspond to HWLOC_MEMBIND_FIRSTTOUCH (Linux, FreeBSD), or HWLOC_MEMBIND_BIND (AIX, HP-UX, Solaris, Windows). This policy is never returned by get membind functions. The nodeset argument is ignored.
HWLOC_MEMBIND_FIRSTTOUCH	Allocate each memory page individually on the local NUMA node of the thread that touches it. The given nodeset should usually be hwloc_topology_get_topology_nodeset() so that the touching thread may run and allocate on any node in the system. On AIX, if the nodeset is smaller, pages are allocated locally (if the local node is in the nodeset) or from a random non-local node (otherwise).
HWLOC_MEMBIND_BIND	Allocate memory on the specified nodes. The actual behavior may slightly vary between operating systems, especially when (some of) the requested nodes are full. On Linux, by default, the MPOL_PREFERRED_MANY (or MPOL_PREFERRED) policy is used. However, if the hwloc strict flag is also given, the Linux MPOL_BIND policy is rather used.
HWLOC_MEMBIND_INTERLEAVE	Allocate memory on the given nodes in an interleaved / round-robin manner. The precise layout of the memory across multiple NUMA nodes is OS/system specific. Interleaving can be useful when threads distributed across the specified NUMA nodes will all be accessing the whole memory range concurrently, since the interleave will then balance the memory references.
HWLOC_MEMBIND_NEXTTOUCH	For each page bound with this policy, by next time it is touched (and next time only), it is moved from its current location to the local NUMA node of the thread where the memory reference occurred (if it needs to be moved at all).
HWLOC_MEMBIND_MIXED	Returned by get_membind() functions when multiple threads or parts of a memory area have differing memory binding policies. Also returned when binding is unknown because binding hooks are empty when the topology is loaded from XML without HWLOC_THISSYSTEM=1, etc.

Function Documentation

hwloc_alloc()

void * hwloc_alloc	(	hwloc_topology_t	topology,
		size_t	len
	)

Allocate some memory.

This is equivalent to malloc(), except that it tries to allocate page-aligned memory from the OS.

Returns

a pointer to the allocated area, or NULL on error.

Note

The allocated memory should be freed with hwloc_free().

hwloc_alloc_membind()

void * hwloc_alloc_membind	(	hwloc_topology_t	topology,
		size_t	len,
		hwloc_const_bitmap_t	set,
		hwloc_membind_policy_t	policy,
		int	flags
	)

Allocate some memory on NUMA memory nodes specified by set.

Returns

a pointer to the allocated area.

NULL with errno set to ENOSYS if the action is not supported and HWLOC_MEMBIND_STRICT is given.

NULL with errno set to EXDEV if the binding cannot be enforced and HWLOC_MEMBIND_STRICT is given.

NULL with errno set to ENOMEM if the memory allocation failed even before trying to bind.

If HWLOC_MEMBIND_BYNODESET is specified, set is considered a nodeset. Otherwise it's a cpuset.

Note

The allocated memory should be freed with hwloc_free().

hwloc_alloc_membind_policy()

static void * hwloc_alloc_membind_policy ( hwloc_topology_t  topology, size_t  len, hwloc_const_bitmap_t  set, hwloc_membind_policy_t  policy, int  flags  )

inlinestatic

Allocate some memory on NUMA memory nodes specified by set.

First, try to allocate properly with hwloc_alloc_membind(). On failure, the current process or thread memory binding policy is changed with hwloc_set_membind() before allocating memory. Thus this function works in more cases, at the expense of changing the current state (possibly affecting future allocations that would not specify any policy).

If HWLOC_MEMBIND_BYNODESET is specified, set is considered a nodeset. Otherwise it's a cpuset.

Returns

a pointer to the allocated area, or NULL on error.

hwloc_free()

int hwloc_free	(	hwloc_topology_t	topology,
		void *	addr,
		size_t	len
	)

Free memory that was previously allocated by hwloc_alloc() or hwloc_alloc_membind().

Returns

0 on success, -1 on error.

hwloc_get_area_membind()

int hwloc_get_area_membind	(	hwloc_topology_t	topology,
		const void *	addr,
		size_t	len,
		hwloc_bitmap_t	set,
		hwloc_membind_policy_t *	policy,
		int	flags
	)

Query the CPUs near the physical NUMA node(s) and binding policy of the memory identified by (addr, len ).

The bitmap set (previously allocated by the caller) is filled with the memory area binding.

This function has two output parameters: set and policy. The values returned in these parameters depend on both the flags passed in and the memory binding policies and nodesets of the pages in the address range.

If HWLOC_MEMBIND_STRICT is specified, the target pages are first checked to see if they all have the same memory binding policy and nodeset. If they do not, -1 is returned and errno is set to EXDEV. If they are identical across all pages, the set and policy are returned in set and policy, respectively.

If HWLOC_MEMBIND_STRICT is not specified, the union of all NUMA node(s) containing pages in the address range is calculated. If all pages in the target have the same policy, it is returned in policy. Otherwise, policy is set to HWLOC_MEMBIND_MIXED.

If HWLOC_MEMBIND_BYNODESET is specified, set is considered a nodeset. Otherwise it's a cpuset.

If any other flags are specified, -1 is returned and errno is set to EINVAL.

Returns

0 on success.

-1 with errno set to EINVAL if len is 0.

hwloc_get_area_memlocation()

int hwloc_get_area_memlocation	(	hwloc_topology_t	topology,
		const void *	addr,
		size_t	len,
		hwloc_bitmap_t	set,
		int	flags
	)

Get the NUMA nodes where memory identified by (addr, len ) is physically allocated.

The bitmap set (previously allocated by the caller) is filled according to the NUMA nodes where the memory area pages are physically allocated. If no page is actually allocated yet, set may be empty.

If pages spread to multiple nodes, it is not specified whether they spread equitably, or whether most of them are on a single node, etc.

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

If HWLOC_MEMBIND_BYNODESET is specified in flags, set is considered a nodeset. Otherwise it's a cpuset.

If len is 0, set is emptied.

Returns

0 on success, -1 on error.

hwloc_get_membind()

int hwloc_get_membind	(	hwloc_topology_t	topology,
		hwloc_bitmap_t	set,
		hwloc_membind_policy_t *	policy,
		int	flags
	)

Query the default memory binding policy and physical locality of the current process or thread.

The bitmap set (previously allocated by the caller) is filled with the process or thread memory binding.

This function has two output parameters: set and policy. The values returned in these parameters depend on both the flags passed in and the current memory binding policies and nodesets in the queried target.

Passing the HWLOC_MEMBIND_PROCESS flag specifies that the query target is the current policies and nodesets for all the threads in the current process. Passing HWLOC_MEMBIND_THREAD specifies that the query target is the current policy and nodeset for only the thread invoking this function.

If neither of these flags are passed (which is the most portable method), the process is assumed to be single threaded. This allows hwloc to use either process-based OS functions or thread-based OS functions, depending on which are available.

HWLOC_MEMBIND_STRICT is only meaningful when HWLOC_MEMBIND_PROCESS is also specified. In this case, hwloc will check the default memory policies and nodesets for all threads in the process. If they are not identical, -1 is returned and errno is set to EXDEV. If they are identical, the values are returned in set and policy.

Otherwise, if HWLOC_MEMBIND_PROCESS is specified (and HWLOC_MEMBIND_STRICT is not specified), the default set from each thread is logically OR'ed together. If all threads' default policies are the same, policy is set to that policy. If they are different, policy is set to HWLOC_MEMBIND_MIXED.

In the HWLOC_MEMBIND_THREAD case (or when neither HWLOC_MEMBIND_PROCESS or HWLOC_MEMBIND_THREAD is specified), there is only one set and policy; they are returned in set and policy, respectively.

If HWLOC_MEMBIND_BYNODESET is specified, set is considered a nodeset. Otherwise it's a cpuset.

If any other flags are specified, -1 is returned and errno is set to EINVAL.

Returns

0 on success, -1 on error.

hwloc_get_proc_membind()

int hwloc_get_proc_membind	(	hwloc_topology_t	topology,
		hwloc_pid_t	pid,
		hwloc_bitmap_t	set,
		hwloc_membind_policy_t *	policy,
		int	flags
	)

Query the default memory binding policy and physical locality of the specified process.

The bitmap set (previously allocated by the caller) is filled with the process memory binding.

This function has two output parameters: set and policy. The values returned in these parameters depend on both the flags passed in and the current memory binding policies and nodesets in the queried target.

Passing the HWLOC_MEMBIND_PROCESS flag specifies that the query target is the current policies and nodesets for all the threads in the specified process. If HWLOC_MEMBIND_PROCESS is not specified (which is the most portable method), the process is assumed to be single threaded. This allows hwloc to use either process-based OS functions or thread-based OS functions, depending on which are available.

Note that it does not make sense to pass HWLOC_MEMBIND_THREAD to this function.

If HWLOC_MEMBIND_STRICT is specified, hwloc will check the default memory policies and nodesets for all threads in the specified process. If they are not identical, -1 is returned and errno is set to EXDEV. If they are identical, the values are returned in set and policy.

Otherwise, set is set to the logical OR of all threads' default set. If all threads' default policies are the same, policy is set to that policy. If they are different, policy is set to HWLOC_MEMBIND_MIXED.

If HWLOC_MEMBIND_BYNODESET is specified, set is considered a nodeset. Otherwise it's a cpuset.

If any other flags are specified, -1 is returned and errno is set to EINVAL.

Returns

0 on success, -1 on error.

Note

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

hwloc_set_area_membind()

int hwloc_set_area_membind	(	hwloc_topology_t	topology,
		const void *	addr,
		size_t	len,
		hwloc_const_bitmap_t	set,
		hwloc_membind_policy_t	policy,
		int	flags
	)

Bind the already-allocated memory identified by (addr, len) to the NUMA node(s) specified by set.

If HWLOC_MEMBIND_BYNODESET is specified, set is considered a nodeset. Otherwise it's a cpuset.

Returns

0 on success or if len is 0.

-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_membind()

int hwloc_set_membind	(	hwloc_topology_t	topology,
		hwloc_const_bitmap_t	set,
		hwloc_membind_policy_t	policy,
		int	flags
	)

Set the default memory binding policy of the current process or thread to prefer the NUMA node(s) specified by set.

If neither HWLOC_MEMBIND_PROCESS nor HWLOC_MEMBIND_THREAD is specified, the current process is assumed to be single-threaded. This is the most portable form as it permits hwloc to use either process-based OS functions or thread-based OS functions, depending on which are available.

If HWLOC_MEMBIND_BYNODESET is specified, set is considered a nodeset. Otherwise it's a cpuset.

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_membind()

int hwloc_set_proc_membind	(	hwloc_topology_t	topology,
		hwloc_pid_t	pid,
		hwloc_const_bitmap_t	set,
		hwloc_membind_policy_t	policy,
		int	flags
	)

Set the default memory binding policy of the specified process to prefer the NUMA node(s) specified by set.

If HWLOC_MEMBIND_BYNODESET is specified, set is considered a nodeset. Otherwise it's a cpuset.

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.

Note

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