numa_types man page on DigitalUNIX

Printed from http://www.polarhome.com/service/man/?qf=numa_types&af=0&tf=2&of=DigitalUNIX

numa_types(4)							 numa_types(4)

NAME
       numa_types - Data types used by NUMA application interfaces

SYNOPSIS
       #include <sys/numa_types.h>

DESCRIPTION
       This  reference	page lists and describes the data types, flags, struc‐
       tures, and unions that are defined in the <numa_types.h> header file to
       support the HP Tru64 UNIX NUMA APIs.

       The  program must call radsetcreate() or cpusetcreate() to allocate the
       RAD set or CPU set associated with any fields defined as type  radset_t
       or cpuset_t, respectively.

       Note  that  numa_types.h	 is indirectly included by the <numa.h> header
       file, which is the header file more frequently specified in the	SYNOP‐
       SIS sections of reference pages for NUMA-related functions.

   Definitions
       The <numa_types> header defines the following data types, flags, struc‐
       tures, and unions, and associated symbolic values: Reserved for	future
       use.   A	 structure type that defines the policy and associated parame‐
       ters for memory allocation.  The memalloc_attr structures  are  associ‐
       ated with memory objects and with processes and threads. This structure
       contains the following members: Specifies the memory allocation	policy
       (as described in the entry for memalloc_policy_t). Remaining members of
       the structure contain parameters used to implement this policy.	Speci‐
       fies  the primary or preferred RAD (region) from which to allocate mem‐
       ory for the MPOL_DIRECTED memory allocation policy.  Specifies the dis‐
       tance  to  overflow.  This  value  is not currently used for any memory
       allocation  policy.   Specifies	the  stride   (in   pages)   for   the
       MPOL_STRIPED  memory  allocation	 policy.   Specifies  the page size in
       bytes. This value is not currently used for any memory allocation  pol‐
       icy.   If  mattr_policy	is MPOL_DIRECTED or MPOL_THREAD, specifies the
       overflow RAD set.

	      If mattr_policy is MPOL_STRIPED, specifies the  RAD  set	across
	      which  memory  is striped.  An enumeration type that determines,
	      along with associated parameter attributes, how memory  will  be
	      allocated	 for  a	 memory	 object	 or a kernel memory allocation
	      request.	Valid policy values are: Allocate pages from a	speci‐
	      fied  (meaning  preferred)  RAD  with overflow into a specified,
	      possibly NULL, overflow RAD set.	 Equivalent  to	 MPOL_DIRECTED
	      but  having  the thread context determine the preferred RAD from
	      which pages are allocated. (In this case, the mattr_rad value is
	      ignored.)	  Allocate  pages  so  that  they are striped across a
	      specified RAD set by using a specified (page multiple) stripe as
	      specified	 by  the  mattr_stride	member	of the memalloc_attr_t
	      structure. Starting with a specified RAD, pages  will  be	 allo‐
	      cated  from  RADs in the RAD set in order of increasing RAD num‐
	      ber. An mattr_stride number of pages will be allocated from each
	      RAD  before  pages  are allocated from the next RAD. After pages
	      are allocated from the highest numbered RAD in the set,  alloca‐
	      tion  will  wrap, which means that pages are next allocated from
	      the lowest numbered RAD in the set.  Replicate pages on the home
	      RAD of the thread that caused the pages to be allocated.

					    Note

	      MPOL_REPLICATED is the default and only memory allocation policy
	      supported for shared library text pages and program text	pages.
	      Furthermore,  the	 operating system ignores MPOL_REPLICATED when
	      it is specified for other types  of  memory  (stack,  heap,  and
	      shared  data  pages).  Therefore, specifying MPOL_REPLICATED has
	      no effect in the current implementation.

	      The following modifier may be combined (by using	a  logical  OR
	      operation) with the preceding policies: Disable automatic migra‐
	      tion of pages by the system.

	      The following table indicates how different memory policies  are
	      supported	 for  the  different  types  of	 memory in the program
	      address space:

	      SHPT = shared program text
	      SHLT = shared library text
	      SVSHM = System V shared memory
	      STACK = program or thread stack
	      DATA = initialized program data
	      BSS = uninitialized program data and heap
	      MPRIV = [n]mmap(MAP_PRIVATE)
	      MSHAR = [n]mmap(MAP_SHARED)
	      MANON = [n]mmap(MAP_ANONYMOUS)

	      ──────────────────────────────────────────────────────────────────────────
			MPOL_DIRECTED	MPOL_THREAD	 MPOL_STRIPED	 MPOL_REPLI‐
									 CATED
	      ──────────────────────────────────────────────────────────────────────────
	      SHPT	Ignored *	Ignored *	 Ignored *	 Default

	      SHLT	Ignored *	Ignored *	 Ignored *	 Default

	      SVSHM	As specified.	As   specified	 Default ***	 =MPOL_DIRECTED
					but no	migra‐
					tion. **

	      STACK	As specified.	Default		 As specified.	 =MPOL_DIRECTED

	      DATA	As specified.	Default		 As specified	 =MPOL_DIRECTED

	      BSS	As specified.	Default		 As specified.	 =MPOL_DIRECTED

	      MPRIV	As specified.	Default		 As specified.	 =MPOL_DIRECTED

	      MSHAR	As specified.	Default		 As specified.	 =MPOL_DIRECTED

	      MANON	As specified.	Default		 As specified.	 =MPOL_DIRECTED

	      ──────────────────────────────────────────────────────────────────────────

	      “Ignored” means only that the MPOL_* flag is ignored; the	 nmad‐
	      vise()  call  is	still  checked	for  errors and any applicable
	      behaviors are performed.	When thread-local memory allocation is
	      specified	 for  System  V	 shared memory segments, the policy is
	      used only for pages that have not yet been allocated, have  been
	      paged out, or have been discarded (by means of the MADV_DONTNEED
	      flag on an nmadvise() call). Existing  pages  are	 not  migrated
	      according	 to the new memory allocation policy in order to avoid
	      thrashing; in other words, the MADV_CURRENT flag on a nmadvise()
	      call is treated as MADV_DONTNEED with respect to System V shared
	      memory segments.	The default memory allocation policy for  Sys‐
	      tem  V  shared memory segments is configurable. If the shm_allo‐
	      cate_striped attribute of the IPC kernel subsystem is set	 to  1
	      (the  default), SVSHM segments are striped. If this attribute is
	      set to 0, SVSHM segments are allocated by using the thread local
	      policy (MPOL_THREAD).  Structures of this type are supported for
	      internal use only.   Identifier  for  a  NUMA  Scheduling	 Group
	      (NSG).  A structure that describes the NUMA topology attributes.
	      This structure is used with the nloc()  function	to  query  the
	      NUMA system topology or with the nfork() function to specify the
	      set of RADs from which a RAD will be selected for a new process.
	      This  structure  contains	 the  following	 members:  The type of
	      resource (as described in the entry for type  rsrctype_t).   The
	      resource	descriptor,  which  identifies	the  instance  of  the
	      resource specified by the nattr_type value. Specify  this	 field
	      as  a union with the appropriate resource handle as described in
	      the entry for rsrcdescr_t. For example,  a  resource  descriptor
	      for a RAD set might be specified as:

	      nat.nattr_descr.rd_radset

	      For  this	 descriptor,  nat  is a structure of type numa_attr_t,
	      rd_radset is the resource handle, and  nat.nattr_type  has  been
	      set  to  R_RAD.	The  distance  to  the requested resource. The
	      nloc() function returns a set of RADs  that  are	less  than  or
	      equal  this distance from the specified resource.	 Symbolic val‐
	      ues used by certain functions. The nloc(3) and  nfork(3)	refer‐
	      ence  pages  list	 and  describe these values.  Identifier for a
	      Resource Affinity Domain (RAD), which is	a  grouping  of	 basic
	      system resources that form a NUMA building block. NUMA platforms
	      contain multiple RADs, and each RAD can  contain	zero  or  more
	      CPUs, memory arrays, and I/O busses. Single-processor and multi‐
	      processor platforms  that	 do  not  use  NUMA  architecture  are
	      treated by the operating system as single-RAD systems.

	      The  radid_t  data  type	is a generic, integral type, for which
	      there is the following symbolic value: No valid  RAD  ID.	 Func‐
	      tions  return  RAD_NONE when no RAD matches the specified crite‐
	      ria, there are no more RADs in a RAD set, and so forth.	A  set
	      of RADs. This type is used to specify a set of radid_t values to
	      the NUMA APIs. A subset of these APIs perform  operations	 on  a
	      set  of  RADs  and  manage radset_t as an opaque type.  A opaque
	      type used in an enumeration or an iteration operation on	a  RAD
	      set.  This type stores the current cursor position during a scan
	      of the members in a RAD set. See rad_foreach(3) for more	infor‐
	      mation.	A structure returned by the rad_get_info() function to
	      describe the state and resources associated  with	 a  RAD.  This
	      structure	 contains the following members: The RAD revision num‐
	      ber.  The RAD identifier.	  The  RAD  state,  whose  values  are
	      listed  in  the description of the rad_state_t type.  The amount
	      of physical memory present in the RAD.  The  current  amount  of
	      free  memory  available  in the RAD.  The set of CPUs associated
	      with the RAD.

	      See rad_get_info(3) for  more  information  about	 querying  RAD
	      information.   A	RAD's  software state. The defined states are:
	      The specified RAD exists and is on line. Processes  and  threads
	      may  be assigned to and memory allocated to the RAD.  The speci‐
	      fied RAD exists but is not currently on line.  No	 processes  or
	      threads  may  be assigned to, nor memory allocated to, this RAD;
	      however, its resource complement may be queried.

					    Note

	      RAD_ONLINE is the only state currently supported.	  An  enumera‐
	      tion  type that specifies the kind of resource with which affin‐
	      ity is desired. Functions that perform NUMA topology queries and
	      resource	binding	 pass  rsrctype_t arguments along with rsrcde‐
	      scr_t descriptors to identify resources. The following  symbolic
	      values  identify	the type of resource being specified by a par‐
	      ticular descriptor: A RAD set.  A file or device	referenced  by
	      an  open file descriptor.	 A file or device specified by a path‐
	      name.  A System V shared memory segment that is referenced by  a
	      shared memory ID.	 A process that is referenced by a pid_t iden‐
	      tifier.  A physical memory mapped region that is referenced by a
	      process virtual address. This resource type is used to find RADs
	      that are equal to or less than a specified distance from a  par‐
	      ticular  memory  location. See nloc(3).  A NUMA Scheduling Group
	      that is referenced by an nsgid_t identifier.   A	union  of  the
	      various  resource handles, or descriptors, that are specified by
	      different resource type (rsrctype_t) values. Along with an rsrc‐
	      type_t  argument,	 an  rsrcdescr_t  handle  is  included	on the
	      nattr_descr argument to the NUMA APIs that perform NUMA topology
	      queries and resource binding. The resource handles for different
	      resource types are as follows: If the rsrctype_t value is R_RAD,
	      a	 RAD  set  is the resource, for which rd_radset is the handle.
	      If the rsrctype_t value is R_FILDES,  an	open  file  or	device
	      (referenced  by  descriptor) is the resource, for which rd_fd is
	      the handle.  If the rsrctype_t value is R_PATH, a file or device
	      (referenced  by pathname) is the resource, for which rd_pathname
	      is the handle.  If the rsrctype_t value is  R_SHM,  a  System  V
	      shared memory segment is the resource, for which rd_shmid is the
	      handle.  If the rsrctype_t value is  R_PID,  a  process  is  the
	      resource,	 for  which  rd_pid  is the handle.  If the rsrctype_t
	      value is R_MEM,  a  mapped  region  of  virtual  memory  is  the
	      resource,	 for  which  rd_addr is the handle.  If the rsrctype_t
	      value is R_NSG, a NUMA Scheduling Group  is  the	resource,  for
	      which rd_nsg is the handle.

	      See nloc(3) and nfork(3) for information about using rsrcdescr_t
	      handles to query and bind resources in the context of NUMA  sys‐
	      tem topology.  A structure that specifies the access permissions
	      and associated parameters and statistics for a  NUMA  Scheduling
	      Group  (NSG).  This  structure contains the following members: A
	      subordinate structure that contains the NSG access  permissions.
	      The  number  of processes attached to the NSG.  A structure that
	      specifies a thread (process ID and thread	 ID)  attached	to  an
	      NSG.  This structure contains the following members: The process
	      ID of a thread that is  attached	to  an	NSG.   The  thread  ID
	      (index) of a thread that is attached to an NSG.

SEE ALSO
       Functions:    nfork(3),	  nloc(3),    numa_intro(3),   rad_foreach(3),
       rad_get_info(3)

								 numa_types(4)
[top]

List of man pages available for DigitalUNIX

Copyright (c) for man pages and the logo by the respective OS vendor.

For those who want to learn more, the polarhome community provides shell access and support.

[legal] [privacy] [GNU] [policy] [cookies] [netiquette] [sponsors] [FAQ]
Tweet
Polarhome, production since 1999.
Member of Polarhome portal.
Based on Fawad Halim's script.
....................................................................
Vote for polarhome
Free Shell Accounts :: the biggest list on the net