DistDataCollector Class Reference

DistDataCollector is a class that collects data stored on multiple MPI processes into a large array stored on a designated root rank. More...

#include <DistDataCollector.h>

Public Member Functions
	DistDataCollector (MPI_Comm comm, int numChunks, int numSize, int rootRank=0)
	Constructor.
	~DistDataCollector ()
	Destructor.
MPI_Win &	getWin ()
void	fence ()
void	startEpoch ()
	Start an epoch for RMA operations.
void	flush ()
	Ensure that the RMA operation is completed and the data are visible to the manager.
void	endEpoch ()
	End an epoch for RMA operations.
void	put (int chunkId, const double *data)
	Put the local data into the collected array.
void	putAsync (int chunkId, const double *data)
	Put the local data into the collected array (non-blocking).
void	accumulate (int chunkId, const double *data)
	Atomically accumulate (add) local data into a chunk on rootRank. Uses MPI_Accumulate with MPI_SUM under a shared lock, so concurrent calls from different origins are safe. Blocks until the operation is complete and the result is visible at rootRank.
std::vector< double >	get (int chunkId)
	Get a slice of the remote, collected array to the local worker.
void	get (int chunkId, double *buffer)
	Get a slice of the remote, collected array to the local worker.
void	getAsync (int chunkId, double *buffer)
	Get a slice of the remote, collected array to the local worker (non-blocking).
double *	getCollectedDataPtr ()
int	getNumChunks () const
int	getNumSize () const
void	free ()
	Free the MPI window and empty the collected data.
	DistDataCollector (const DistDataCollector &)=delete
DistDataCollector &	operator= (const DistDataCollector &)=delete
	DistDataCollector (DistDataCollector &&other) noexcept
DistDataCollector &	operator= (DistDataCollector &&other) noexcept

Public Attributes
const double	BAD_VALUE = std::numeric_limits<double>::quiet_NaN()

Detailed Description

DistDataCollector is a class that collects data stored on multiple MPI processes into a large array stored on a designated root rank.

Constructor & Destructor Documentation

DistDataCollector()

DistDataCollector::DistDataCollector	(	MPI_Comm	comm,
		int	numChunks,
		int	numSize,
		int	rootRank = 0 )

Constructor.

Parameters

comm	MPI communicator to use for communication
numChunks	The number of array slices on rootRank
numSize	The size of each slice
rootRank	MPI rank that holds the collected data (default: 0)

Member Function Documentation

accumulate()

void DistDataCollector::accumulate	(	int	chunkId,
		const double *	data )

Atomically accumulate (add) local data into a chunk on rootRank. Uses MPI_Accumulate with MPI_SUM under a shared lock, so concurrent calls from different origins are safe. Blocks until the operation is complete and the result is visible at rootRank.

Parameters

chunkId	Leading index in the collected array
data	Pointer to the values to add (must have numSize elements)

fence()

void DistDataCollector::fence

(

)

Active target synchronization

Note

call this before and after a put operation to ensure that the target data are visible. All ranks must call this method.

get() [1/2]

std::vector< double > DistDataCollector::get

(

int

chunkId

)

Get a slice of the remote, collected array to the local worker.

Parameters

chunkId

Leading index in the collected array

Returns

data array

Note

this should be executed on the source process, typically by the worker

get() [2/2]

void DistDataCollector::get	(	int	chunkId,
		double *	buffer )

Get a slice of the remote, collected array to the local worker.

Parameters

chunkId	Leading index in the collected array
buffer	will hold the fetched data

Note

this should be executed on the source process, typically by the worker

getAsync()

void DistDataCollector::getAsync ( int chunkId, double * buffer )

inline

Get a slice of the remote, collected array to the local worker (non-blocking).

Parameters

chunkId	Leading index in the collected array
buffer	will hold the fetched data

Note

this should be executed on the source process, typically by the worker. This is a non-blocking call which relies on startEpoch/flush/endEpoch to complete

getCollectedDataPtr()

double * DistDataCollector::getCollectedDataPtr ( )

inline

Get the pointer to the collected data

Returns

pointer

Note

this returns a null pointer on ranks other than rootRank

getNumChunks()

int DistDataCollector::getNumChunks ( ) const

inline

Get the number of chunks

Returns

number

getNumSize()

int DistDataCollector::getNumSize ( ) const

inline

Get the size of each chunk

Returns

number

getWin()

MPI_Win & DistDataCollector::getWin ( )

inline

Get the MPI window

put()

void DistDataCollector::put	(	int	chunkId,
		const double *	data )

Put the local data into the collected array.

Parameters

chunkId	Leading index in the collected array
data	Pointer to the local data to inject

Note

this should be executed on the source process, typically by the worker

putAsync()

void DistDataCollector::putAsync ( int chunkId, const double * data )

inline

Put the local data into the collected array (non-blocking).

Parameters

chunkId	Leading index in the collected array
data	Pointer to the local data to inject

Note

this should be executed on the source process, typically by the worker This is a non-blocking call which relies on startEpoch/flush/endEpoch to

---

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

• src/DistDataCollector.h
• src/DistDataCollector.cpp