numa.hpp File Reference

NUMA topology detection: distance/preference matrices and per-domain delegate-thread selection, identifying threads by OpenMP thread id and using sched_getcpu only for the NUMA domain lookup. More...

#include <cstdlib>
#include <jaffarCommon/exceptions.hpp>
#include <jaffarCommon/parallel.hpp>
#include <map>
#include <mutex>
#include <numa.h>
#include <vector>

Go to the source code of this file.

Functions
void	jaffarPlus::initializeNUMA ()
	Initializes NUMA / core-affinity state.

Variables
static int	jaffarPlus::_numaCount
	Number of NUMA domains detected.
static int	jaffarPlus::_threadCount
	Number of threads to use.
static std::vector< std::vector< size_t > >	jaffarPlus::_numaDistanceMatrix
	NUMA distance matrix; entry [i][j] is the reported distance from domain i to domain j.
static std::vector< std::vector< size_t > >	jaffarPlus::_numaPreferenceMatrix
	NUMA preference matrix; row i lists domains ordered by ascending distance from domain i (ties rotated per source domain).
static std::vector< ssize_t >	jaffarPlus::_numaDelegateThreadId
	Delegate thread (OpenMP thread id) per NUMA domain, or -1 if none assigned.
static thread_local int	jaffarPlus::_preferredNumaDomain
	Thread-local preferred NUMA domain (the NUMA node of the CPU the thread currently runs on).
static thread_local int	jaffarPlus::_myThreadId
	Thread-local dense OpenMP thread id of the current thread.

Detailed Description

NUMA topology detection: distance/preference matrices and per-domain delegate-thread selection, identifying threads by OpenMP thread id and using sched_getcpu only for the NUMA domain lookup.

Definition in file numa.hpp.

Function Documentation

initializeNUMA()

void jaffarPlus::initializeNUMA

(

)

Initializes NUMA / core-affinity state.

Detects the thread count and NUMA domain count, optionally overrides the thread count from the JAFFAR_ENGINE_OVERRIDE_MAX_THREAD_COUNT environment variable, builds the NUMA distance and preference matrices, and selects one delegate thread per NUMA domain. Each worker identifies itself by its OpenMP thread id and derives its preferred NUMA domain from sched_getcpu.

Exceptions

A	runtime error if the system does not provide NUMA detection support.

Definition at line 44 of file numa.hpp.

Variable Documentation

_myThreadId

thread_local int jaffarPlus::_myThreadId

static

Thread-local dense OpenMP thread id of the current thread.

Definition at line 34 of file numa.hpp.

_numaCount

int jaffarPlus::_numaCount

static

Number of NUMA domains detected.

Definition at line 21 of file numa.hpp.

_numaDelegateThreadId

std::vector<ssize_t> jaffarPlus::_numaDelegateThreadId

static

Delegate thread (OpenMP thread id) per NUMA domain, or -1 if none assigned.

Definition at line 30 of file numa.hpp.

_numaDistanceMatrix

std::vector<std::vector<size_t> > jaffarPlus::_numaDistanceMatrix

static

NUMA distance matrix; entry [i][j] is the reported distance from domain i to domain j.

Definition at line 25 of file numa.hpp.

_numaPreferenceMatrix

std::vector<std::vector<size_t> > jaffarPlus::_numaPreferenceMatrix

static

NUMA preference matrix; row i lists domains ordered by ascending distance from domain i (ties rotated per source domain).

Definition at line 28 of file numa.hpp.

_preferredNumaDomain

thread_local int jaffarPlus::_preferredNumaDomain

static

Thread-local preferred NUMA domain (the NUMA node of the CPU the thread currently runs on).

Definition at line 32 of file numa.hpp.

_threadCount

int jaffarPlus::_threadCount

static

Number of threads to use.

Definition at line 23 of file numa.hpp.