thread_binding.h File Reference

Startup check that the MPI launcher did not pin all threads of a rank onto too few hardware threads. More...

#include <madness/world/madness_exception.h>
#include <madness/world/safempi.h>
#include <memory>
#include <string>
#include <utility>
#include <vector>

Include dependency graph for thread_binding.h:

This graph shows which files directly or indirectly include this file:

Go to the source code of this file.

Classes
class	madness::ThreadBindingException
	Thrown by check_thread_binding() when the CPU binding cannot support the thread pool. More...
struct	madness::ThreadBindingInfo
	Description of the CPU affinity of the calling process. More...
struct	madness::detail::ThreadBindingMessage
	Owns the (long) diagnostic text of a ThreadBindingException. More...

Namespaces
namespace	madness
	Namespace for all elements and tools of MADNESS.
namespace	madness::detail

Functions
void	madness::check_thread_binding (const SafeMPI::Intracomm &comm, int nthread_app)
	Throw if the CPU binding cannot accommodate the thread pool.
std::string	madness::cpu_list_to_string (const std::vector< int > &cpus)
	Render a list of cpu ids in compact form, e.g. {0,1,2,3,8} -> "0-3,8".
ThreadBindingInfo	madness::get_thread_binding_info ()
	Query the CPU affinity mask and NUMA placement of the calling process.

Detailed Description

Startup check that the MPI launcher did not pin all threads of a rank onto too few hardware threads.

Most MPI launchers bind a rank to a single core by default (e.g. Open MPI's --bind-to core when there are more cores than ranks on a node). A multi-threaded MADNESS rank inherits that affinity mask for every thread it spawns, so the whole rank time-shares one hardware thread. The job still runs and produces correct results, but one to two orders of magnitude slower, which is easy to miss. check_thread_binding() detects this at startup and throws a ThreadBindingException carrying an actionable error message.