mpi::OpenMPI Class Reference

The MPI singleton is responsible for data exchange between compute nodes via OpenMPI. More...

#include <openmpi.h>

Public Member Functions
int	getRank () const
	returns the rank of the node. More...
int	getSize () const
	returns the total number of nodes. More...
void	send (int target, int source, math::CFunction &f)
	sends a copy of the math::CFunction from node source to node target. More...
void	send (int target, int source, double &v)
	sends a copy of double value v from node source to node target. More...
void	combine (math::CFunction &f)
	assembles the partial information of the individual compute nodes to a full copy on each one. More...
void	sync (math::CFunction &f, int master)
	distributes the math::CFunction instance f on node getRank()==master to all other nodes. More...
void	sync (double &v, int master)
	distributes the double value v on node getRank()==master to all other nodes. More...
void	sync ()
	blocks until all nodes have reached this function call.

Static Public Member Functions
static OpenMPI &	getInstance ()
	returns the only instance of the OpenMPI class. More...

Protected Attributes
int	m_rank
int	m_size

Static Protected Attributes
static std::auto_ptr< OpenMPI >	m_ptr = std::auto_ptr<OpenMPI>(0)

Friends
class	std::auto_ptr< OpenMPI >

Detailed Description

The MPI singleton is responsible for data exchange between compute nodes via OpenMPI.

First and foremost, the MPI class provides convenience functions for the exchange of math::CFunction instances. Secondly, combine allows to reassemble a function which has been jointly calculated on the participating compute nodes. Third, it provides barrier functions to synchronize the nodes.

Member Function Documentation

void mpi::OpenMPI::combine

(

math::CFunction &

f

)

assembles the partial information of the individual compute nodes to a full copy on each one.

Typically (in this program), the calculation of discretized functions \(f(x_i)\) is parallelized by assigning the individual calculations for different arguments \(x_i\) in a round-robin fashion among the compute nodes. Therefore compute node \(p\) holds the data points \( f(x_p), f(x_{p+N}),f(x_{p+2N}), \dots \) with \(N\) the total number of compute nodes. This function reassembles the complete function by collecting all partial information on the master node, rebuilding the full function and distributing it to all nodes. Function blocks until operation is complete.

Parameters

[in,out]

f

data contribution (in) / storage for assembled data (out)

OpenMPI & mpi::OpenMPI::getInstance ( )

static

returns the only instance of the OpenMPI class.

If OpenMPI has not been accessed previously an instance is created. The auto_ptr ensures that the object is properly destructed when the the application closes. OpenMPI is initialized when the object is first constructed.

Returns

the only instance of the OpenMPI class

int mpi::OpenMPI::getRank ( ) const

inline

returns the rank of the node.

Returns

rank of the node

int mpi::OpenMPI::getSize ( ) const

inline

returns the total number of nodes.

Returns

total number of nodes.

void mpi::OpenMPI::send	(	int	target,
		int	source,
		math::CFunction &	f
	)

sends a copy of the math::CFunction from node source to node target.

If the return value of getRank() is neither equal to source nor to target the function does nothing. For getRank()==source a copy of the math::CFunction instance referenced by f is send to node target. The contents of the math::CFunction instance referenced by f on target is replaced with the received data. Function blocks until operation is complete.

Parameters

[in]	target	target node
[in]	source	source node
[in,out]	f	data to send (source) / storage location (target)

void mpi::OpenMPI::send	(	int	target,
		int	source,
		double &	v
	)

sends a copy of double value v from node source to node target.

If the return value of getRank() is neither equal to source nor to target the function does nothing. For getRank()==source the double value v] is send to node target. The contents of the double value referenced by f on target is replaced with the received data. Function blocks until operation is complete.

Parameters

[in]	target	target node
[in]	source	source node
[in,out]	v	data to send (source) / storage location (target)

void mpi::OpenMPI::sync	(	math::CFunction &	f,
		int	master
	)

distributes the math::CFunction instance f on node getRank()==master to all other nodes.

Function blocks until operation is complete.

Parameters

[in,out]	f	data to send (master) / storage location (others)
[in]	master	node with master copy

void mpi::OpenMPI::sync	(	double &	v,
		int	master
	)

distributes the double value v on node getRank()==master to all other nodes.

Function blocks until operation is complete.

Parameters

[in,out]	f	data to send (master) / storage location (others)
[in]	master	node with master copy

---

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

• /home/chris/Development/Release/NanoStructures/mpi/openmpi.h
• /home/chris/Development/Release/NanoStructures/mpi/openmpi.cpp