grape::cuda::GPUMessageManager Class Reference

Public Member Functions
void	Init (const grape::CommSpec &comm_spec)
	Initialize message manager.
void	InitBuffer (size_t send_buffer_capacity, size_t recv_buffer_capacity)
void	DropBuffer ()
void	Start ()
	This function will be called before Init step of applications.
void	StartARound ()
	This function will be called before each evaluation step of applications.
void	FinishARound ()
	This function will be called after each evaluation step of applications.
void	Finalize () const
	This function will be called after the evaluation of applications.
bool	ToTerminate () const
	This function will be called by worker after a step to determine whether evaluation is terminated.
size_t	GetMsgSize () const
	Get size of messages sent by this message manager instance. The return value is valid only after FinishARound is called. StartARound will reset the value to zero.
void	ForceContinue ()
	Force continue to evaluate one more round even if all workers stop sending message.
double	GetAccumulatedCommTime () const
dev::MessageManager	DeviceObject ()
void	DeviceWarmup (size_t np)
	For some GPU servers, the first kernel always takes a long time. This is a Dummy function to warm up the device.
template<typename GRAPH_T , typename MESSAGE_T = grape::EmptyType, typename FUNC_T >
void	ParallelProcess (const GRAPH_T &frag, FUNC_T func)
template<typename MESSAGE_T = grape::EmptyType, typename FUNC_T >
void	ParallelProcess (FUNC_T func)
Stream &	stream ()
ncclComm_t	nccl_comm ()

Private Member Functions
bool	syncLengths ()

Private Attributes
InArchiveGroup	to_send_
OutArchiveGroup	to_recv_
pinned_vector< dev::OutArchive >	pinned_to_recv_
thrust::device_vector< dev::InArchive >	d_to_send_
thrust::device_vector< dev::OutArchive >	d_to_recv_
Event	computation_finished_
grape::CommSpec	comm_spec_
std::shared_ptr< ncclComm_t >	nccl_comm_
Stream	compute_stream_
Stream	comm_stream_
std::vector< size_t >	lengths_out_
std::vector< size_t >	lengths_in_
size_t	sent_size_ {}
double	total_memcpy_time_ {}
bool	to_terminate_ {}
bool	force_continue_ {}
fid_t	fid_
fid_t	fnum_

Member Function Documentation

DeviceWarmup()

void grape::cuda::GPUMessageManager::DeviceWarmup ( size_t  np )

inline

For some GPU servers, the first kernel always takes a long time. This is a Dummy function to warm up the device.

This function can be called by applications.

ForceContinue()

void grape::cuda::GPUMessageManager::ForceContinue ( )

inline

Force continue to evaluate one more round even if all workers stop sending message.

This function can be called by applications.

GetMsgSize()

size_t grape::cuda::GPUMessageManager::GetMsgSize ( ) const

inline

Get size of messages sent by this message manager instance. The return value is valid only after FinishARound is called. StartARound will reset the value to zero.

Returns

Size of messages sent by this message manager instance.

Init()

void grape::cuda::GPUMessageManager::Init ( const grape::CommSpec &  comm_spec )

inline

Initialize message manager.

Parameters

comm	MPI_Comm object.

ToTerminate()

bool grape::cuda::GPUMessageManager::ToTerminate ( ) const

inline

This function will be called by worker after a step to determine whether evaluation is terminated.

Returns

Whether evaluation is terminated.