interface of core Allreduce engine More...

#include <engine.h>

Inheritance diagram for rabit::engine::IEngine:

Public Types
typedef void()	PreprocFunction(void *arg)
	Preprocessing function, that is called before AllReduce, used to prepare the data used by AllReduce.

typedef void()	ReduceFunction(const void src, void dst, int count, const MPI::Datatype &dtype)
	reduce function, the same form of MPI reduce function is used, to be compatible with MPI interface In all the functions, the memory is ensured to aligned to 64-bit which means it is OK to cast src,dst to double* int* etc

Public Member Functions
	~IEngine ()=default
	virtual destructor

virtual void	Allgather (void *sendrecvbuf, size_t total_size, size_t slice_begin, size_t slice_end, size_t size_prev_slice)=0
	Allgather function, each node have a segment of data in the ring of sendrecvbuf, the data provided by current node k is [slice_begin, slice_end), the next node's segment must start with slice_end after the call of Allgather, sendrecvbuf_ contains all the contents including all segments use a ring based algorithm.

virtual void	Allreduce (void sendrecvbuf_, size_t type_nbytes, size_t count, ReduceFunction reducer, PreprocFunction prepare_fun=nullptr, void prepare_arg=nullptr)=0
	performs in-place Allreduce, on sendrecvbuf this function is NOT thread-safe

virtual void	Broadcast (void *sendrecvbuf_, size_t size, int root)=0
	broadcasts data from root to every other node

virtual int	LoadCheckPoint ()=0

virtual void	CheckPoint ()=0
	Increase internal version number. Deprecated.

virtual int	VersionNumber () const =0

virtual int	GetRingPrevRank () const =0
	gets rank of previous node in ring topology

virtual int	GetRank () const =0
	gets rank of current node

virtual int	GetWorldSize () const =0
	gets total number of nodes

virtual bool	IsDistributed () const =0
	whether we run in distribted mode

virtual std::string	GetHost () const =0
	gets the host name of the current node

virtual void	TrackerPrint (const std::string &msg)=0
	prints the msg in the tracker, this function can be used to communicate progress information to the user who monitors the tracker

Detailed Description

interface of core Allreduce engine

Member Typedef Documentation

◆ PreprocFunction

typedef void() rabit::engine::IEngine::PreprocFunction(void *arg)

Preprocessing function, that is called before AllReduce, used to prepare the data used by AllReduce.

Parameters

arg	additional possible argument used to invoke the preprocessor

◆ ReduceFunction

typedef void() rabit::engine::IEngine::ReduceFunction(const void *src, void *dst, int count, const MPI::Datatype &dtype)

reduce function, the same form of MPI reduce function is used, to be compatible with MPI interface In all the functions, the memory is ensured to aligned to 64-bit which means it is OK to cast src,dst to double* int* etc

Parameters

src	pointer to source space
dst	pointer to destination reduction
count	total number of elements to be reduced (note this is total number of elements instead of bytes) the definition of the reduce function should be type aware
dtype	the data type object, to be compatible with MPI reduce

Member Function Documentation

◆ Allgather()

virtual void rabit::engine::IEngine::Allgather	(	void *	sendrecvbuf,
		size_t	total_size,
		size_t	slice_begin,
		size_t	slice_end,
		size_t	size_prev_slice
	)

pure virtual

Allgather function, each node have a segment of data in the ring of sendrecvbuf, the data provided by current node k is [slice_begin, slice_end), the next node's segment must start with slice_end after the call of Allgather, sendrecvbuf_ contains all the contents including all segments use a ring based algorithm.

Parameters

sendrecvbuf_	buffer for both sending and receiving data, it is a ring conceptually
total_size	total size of data to be gathered
slice_begin	beginning of the current slice
slice_end	end of the current slice
size_prev_slice	size of the previous slice i.e. slice of node (rank - 1) % world_size

Implemented in rabit::engine::AllreduceMock, rabit::engine::AllreduceBase, and rabit::engine::MPIEngine.

◆ Allreduce()

virtual void rabit::engine::IEngine::Allreduce	(	void *	sendrecvbuf_,
		size_t	type_nbytes,
		size_t	count,
		ReduceFunction	reducer,
		PreprocFunction	prepare_fun = `nullptr`,
		void *	prepare_arg = `nullptr`
	)

pure virtual

performs in-place Allreduce, on sendrecvbuf this function is NOT thread-safe

Parameters

sendrecvbuf_	buffer for both sending and receiving data
type_nbytes	the number of bytes the type has
count	number of elements to be reduced
reducer	reduce function
prepare_func	Lazy preprocessing function, if it is not NULL, prepare_fun(prepare_arg) will be called by the function before performing Allreduce in order to initialize the data in sendrecvbuf. If the result of Allreduce can be recovered directly, then prepare_func will NOT be called
prepare_arg	argument used to pass into the lazy preprocessing function