HaloExchange.hpp File Reference

Asynchronous, split-phase ghost-communication engine. More...

#include "BoxGeometry.hpp"
#include "ghosts/HaloPlan.hpp"
#include "ghosts/particle_packing.hpp"
#include <boost/mpi/request.hpp>
#include <memory>
#include <vector>

Include dependency graph for HaloExchange.hpp:

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

Go to the source code of this file.

Classes
struct	GhostComm::ExchangeBuffers
	Persistent per-neighbor buffer pool for halo exchanges. More...
struct	GhostComm::GhostExchange
	Opaque handle for one in-flight halo exchange. More...

Namespaces
namespace	GhostComm

Functions
GhostExchange	GhostComm::halo_exchange_start (HaloPlan const &plan, BoxGeometry const &box, unsigned data_parts, ExchangeOp op, ExchangeBuffers &bufs)
	Begin a halo exchange using a caller-owned buffer pool.
GhostExchange	GhostComm::halo_exchange_start (HaloPlan const &plan, BoxGeometry const &box, unsigned data_parts, ExchangeOp op)
	Convenience overload that allocates a temporary ExchangeBuffers.
void	GhostComm::halo_exchange_finish (GhostExchange &state)
	Complete a halo exchange: run same-rank copies (overlapping the in-flight messages), wait for all requests, then unpack or reduce.
void	GhostComm::halo_exchange (HaloPlan const &plan, BoxGeometry const &box, unsigned data_parts, ExchangeOp op, ExchangeBuffers &bufs)
	Blocking wrapper using a caller-owned buffer pool (no per-call alloc after warm-up).
void	GhostComm::halo_exchange (HaloPlan const &plan, BoxGeometry const &box, unsigned data_parts, ExchangeOp op)
	Blocking convenience wrapper: start + finish (allocates a temporary ExchangeBuffers; use the pool overload on hot paths).

Detailed Description

Asynchronous, split-phase ghost-communication engine.

Drive a GhostComm::HaloPlan with non-blocking point-to-point MPI (plus same-rank copies), reusing the byte-identical serialization from particle_packing.hpp. The engine is split into a halo_exchange_start phase (post all receives, pack and post all sends) and a halo_exchange_finish phase (same-rank copies, wait, unpack/reduce) so that callers can overlap other work with in-flight messages.

Deadlock-freedom is guaranteed by posting every irecv before any isend and matching each message by (peer, tag); message sizes are known a-priori from the receiving cells' packed size (see GhostComm::calc_transmit_size), except for the PARTNUM bootstrap.

Definition in file HaloExchange.hpp.