windows_topology()
{ // {{{
std::size_t const num_of_cores = hardware_concurrency();
numa_node_numbers_.reserve(num_of_cores);
numa_node_affinity_masks_.reserve(num_of_cores);
ns_numa_node_affinity_masks_.reserve(num_of_cores);
thread_affinity_masks_.reserve(num_of_cores);
ns_thread_affinity_masks_.reserve(num_of_cores);
// Initialize each set of data entirely, as some of the initialization
// routines rely on access to other pieces of topology data. The
// compiler will optimize the loops where possible anyways.
for (std::size_t i = 0; i < num_of_cores; ++i)
numa_node_numbers_.push_back(init_numa_node_number(i));
for (std::size_t i = 0; i < num_of_cores; ++i)
{
numa_node_affinity_masks_.push_back(
init_numa_node_affinity_mask(i, false));
ns_numa_node_affinity_masks_.push_back(
init_numa_node_affinity_mask(i, true));
}
for (std::size_t i = 0; i < num_of_cores; ++i)
{
thread_affinity_masks_.push_back(
init_thread_affinity_mask(i, false));
ns_thread_affinity_masks_.push_back(
init_thread_affinity_mask(i, true));
}
} // }}}
void set_thread_affinity(
std::size_t num_thread
, bool numa_sensitive
, error_code& ec = throws
) const
{
cpu_set_t cpu;
CPU_ZERO(&cpu);
CPU_SET(num_thread % hardware_concurrency(), &cpu);
#if defined(HPX_HAVE_PTHREAD_SETAFFINITY_NP)
if (0 == pthread_setaffinity_np(pthread_self(), sizeof(cpu), &cpu))
sleep(0); // Allow the OS to pick up the change.
#else
if (0 == sched_setaffinity(syscall(SYS_gettid), sizeof(cpu), &cpu))
sleep(0); // Allow the OS to pick up the change.
#endif
else
{
HPX_THROWS_IF(ec, kernel_error
, "hpx::threads::set_thread_affinity"
, "failed to set thread affinity");
}
if (ec)
return;
else if (&ec != &throws)
ec = make_success_code();
}
mask_type init_thread_affinity_mask(
std::size_t num_thread
, bool numa_sensitive
)
{ // {{{
std::size_t num_of_cores = hardware_concurrency();
std::size_t affinity = num_thread % num_of_cores;
ULONG numa_nodes = 1;
if (GetNumaHighestNodeNumber(&numa_nodes))
++numa_nodes;
std::size_t num_of_cores_per_numa_node = num_of_cores / numa_nodes;
ULONGLONG node_affinity_mask = 0;
ULONGLONG mask = 0x01LL;
if (numa_sensitive) {
UCHAR numa_node = UCHAR(affinity % numa_nodes);
if (!GetNumaNodeProcessorMask(numa_node, &node_affinity_mask))
{
HPX_THROW_EXCEPTION(kernel_error
, "hpx::threads::windows_topology::init_thread_affinity_mask"
, boost::str(boost::format(
"failed to initialize thread %1% affinity mask")
% num_thread));
}
mask = least_significant_bit(node_affinity_mask) <<
(affinity / numa_nodes);
}
else {
UCHAR numa_node = UCHAR(get_numa_node_number(num_thread));
if (!GetNumaNodeProcessorMask(numa_node, &node_affinity_mask))
{
HPX_THROW_EXCEPTION(kernel_error
, "hpx::threads::windows_topology::init_thread_affinity_mask"
, boost::str(boost::format(
"failed to initialize thread %1% affinity mask")
% num_thread));
}
mask = least_significant_bit(node_affinity_mask) <<
(affinity % num_of_cores_per_numa_node);
}
while (!(mask & node_affinity_mask)) {
mask <<= 1LL;
if (0 == mask)
mask = 0x01LL;
}
return static_cast<mask_type>(mask);
} // }}}
mask_type get_machine_affinity_mask(
error_code& ec = throws
) const
{
if (&ec != &throws)
ec = make_success_code();
// the machine mask is the bit-or of all masks in the system
std::size_t const num_of_cores = hardware_concurrency();
mask_type mask = 0;
for (std::size_t i = 0; i < num_of_cores; ++i)
mask |= numa_node_affinity_masks_[i];
return mask;
}
mask_type init_numa_node_affinity_mask(
std::size_t num_thread
, bool numa_sensitive
)
{ // {{{
std::size_t num_of_cores = hardware_concurrency();
UCHAR affinity = UCHAR(num_thread % num_of_cores);
ULONG numa_nodes = 1;
if (GetNumaHighestNodeNumber(&numa_nodes))
++numa_nodes;
ULONGLONG mask = 0;
if (numa_sensitive) {
UCHAR numa_node = affinity % numa_nodes;
if (!GetNumaNodeProcessorMask(numa_node, &mask))
{
HPX_THROW_EXCEPTION(kernel_error
, "hpx::threads::windows_topology::init_numa_node_affinity_mask"
, boost::str(boost::format(
"failed to initialize NUMA node affinity mask for "
"thread %1%")
% num_thread));
}
return static_cast<mask_type>(mask);
}
UCHAR numa_node = UCHAR(get_numa_node_number(num_thread));
if (!GetNumaNodeProcessorMask(numa_node, &mask))
{
HPX_THROW_EXCEPTION(kernel_error
, "hpx::threads::windows_topology::init_numa_node_affinity_mask"
, boost::str(boost::format(
"failed to initialize NUMA node affinity mask for "
"thread %1%")
% num_thread));
}
return static_cast<mask_type>(mask);
} // }}}
std::size_t init_numa_node_number(
std::size_t num_thread
)
{ // {{{
if (std::size_t(-1) == num_thread)
return std::size_t(-1);
UCHAR node_number = 0;
if (GetNumaProcessorNode(UCHAR(num_thread), &node_number))
return node_number;
std::size_t num_of_cores = hardware_concurrency();
if (0 == num_of_cores)
num_of_cores = 1; // assume one core
std::size_t num_of_numa_cores = num_of_cores;
ULONG numa_nodes = 0;
if (GetNumaHighestNodeNumber(&numa_nodes) && 0 != numa_nodes)
num_of_numa_cores = num_of_cores / (numa_nodes + 1);
return num_thread / num_of_numa_cores;
} // }}}
