kamping::plugin::DispatchAlltoall< Comm, DefaultContainerType > Class Template Reference

Plugin providing an alltoallv exchange method which calls one of multiple underlying alltoallv exchange algorithms depending on the communication volume. More...

#include <alltoall_dispatch.hpp>

Inheritance diagram for kamping::plugin::DispatchAlltoall< Comm, DefaultContainerType >:

Collaboration diagram for kamping::plugin::DispatchAlltoall< Comm, DefaultContainerType >:

Public Member Functions
template<typename... Args>
auto	alltoallv_dispatch (Args... args) const
	Alltoallv exchange method which uses the communication volume to either exchange the data using GridCommunicator::alltoallv() (latency in about sqrt(comm.size())) or builtin MPI_Alltoallv (potentially linear latency).
void	initialize () const
	Initialized the grid communicator. If not explicitly called by the user this will be done during the first call to alltoallv_dispatch which internally uses grid communication.

Detailed Description

template<typename Comm, template< typename... > typename DefaultContainerType>
class kamping::plugin::DispatchAlltoall< Comm, DefaultContainerType >

Plugin providing an alltoallv exchange method which calls one of multiple underlying alltoallv exchange algorithms depending on the communication volume.

See also

DispatchAlltoall::alltoallv_dispatch() for more information.

Member Function Documentation

alltoallv_dispatch()

template<typename Comm , template< typename... > typename DefaultContainerType>

template<typename... Args>

auto kamping::plugin::DispatchAlltoall< Comm, DefaultContainerType >::alltoallv_dispatch ( Args... args ) const

inline

Alltoallv exchange method which uses the communication volume to either exchange the data using GridCommunicator::alltoallv() (latency in about sqrt(comm.size())) or builtin MPI_Alltoallv (potentially linear latency).

If the bottleneck send communication volume on all ranks is smaller than a given threshold (in number bytes), our grid alltoall communication is used. Otherwise we use the builtin MPI alltoallv exchange.

The following parameters are required:

• kamping::send_buf() containing the data that is sent to each rank. The size of this buffer has to be at least the sum of the send_counts argument.
• kamping::send_counts() containing the number of elements to send to each rank.

The following buffers are optional:

• dispatch_alltoall::comm_volume_threshold() containing the threshold for the maximum bottleneck communication volume in bytes indicating to switch from grid to builtin alltoall exchange. If ommitted, a threshold value of 2000 bytes is used.
• kamping::recv_counts() containing the number of elements to receive from each rank. This parameter is mandatory if kamping::recv_type() is given.
• kamping::recv_buf() containing a buffer for the output. Afterwards, this buffer will contain the data received as specified for send_buf. The buffer will be resized according to the buffer's kamping::BufferResizePolicy. If resize policy is kamping::BufferResizePolicy::no_resize, the buffer's underlying storage must be large enough to store all received elements.

Template Parameters

Args	Automatically deducted template parameters.

Parameters

args	All required and any number of the optional parameters described above.

remove comm_volume_threshold and unpacked send_counts from caller provided argument list before forwarding it underlying allotall exchanges

---

The documentation for this class was generated from the following file:

• include/kamping/plugin/alltoall_dispatch.hpp