Synchronize arrays between processes. More...

#include <CalcServer/MPISync.h>

Inheritance diagram for Aqua::CalcServer::MPISync:

Collaboration diagram for Aqua::CalcServer::MPISync:

Data Structures
class	Exchanger
	Interprocess array synchronization base class. More...

class	Receiver
	Synchronize arrays between processes. More...

class	Sender
	Synchronize arrays between processes. More...

Public Member Functions
	MPISync (const std::string name, const std::string mask, const std::vector< std::string > fields, const std::vector< unsigned int > procs, bool once=false)

	~MPISync ()

void	setup ()

Public Member Functions inherited from Aqua::CalcServer::Tool
	Tool (const std::string tool_name, bool once=false)

virtual	~Tool ()

void	name (const std::string tool_name)

const std::string	name ()

virtual void	setup ()

virtual void	execute ()
	Execute the tool measuring the elapsed time.

virtual Tool *	next_tool ()

size_t	allocatedMemory () const

unsigned int	used_times () const

float	elapsedTime (bool averaged=true) const

float	elapsedTimeVariance () const

float	elapsedTimeDeviation () const

virtual const int	scope_modifier ()

Protected Member Functions
cl_event	_execute (const std::vector< cl_event > events)

Protected Member Functions inherited from Aqua::CalcServer::Tool
int	id_in_pipeline ()

void	next_tool (Tool *tool)

void	allocatedMemory (size_t mem_size)

virtual cl_event	_execute (const std::vector< cl_event > events)

void	addElapsedTime (float elapsed_time)
	Add new data to the average and squared elapsed times.

void	setDependencies (std::vector< std::string > var_names)
	Set the depedencies of the tool.

void	setDependencies (std::vector< InputOutput::Variable * > vars)
	Set the depedencies of the tool.

const std::vector< InputOutput::Variable * >	getDependencies ()
	Get the depedencies of the tool.

Additional Inherited Members
Static Protected Member Functions inherited from Aqua::CalcServer::Tool
static std::vector< cl_kernel >	compile (const std::string source, const std::vector< std::string > names, const std::string flags="")
	Compile an OpenCL source code and generate the corresponding kernel.

static cl_kernel	compile_kernel (const std::string source, const std::string kernel_name, const std::string flags="")
	Compile an OpenCL source code and generate the corresponding kernel.

Detailed Description

Synchronize arrays between processes.

When MPI is enabled, several instances/processes of AQUAgpusph can be simultaneously launched, letting each process compute a subset of the whole simulation, in such a way the global simulation computation can be accelerated.

The data synchronization is a quite expensive operation, both in computational and physical time, so it shall be thoroughly used.

First, the tool should dispose the data in a convenient way, which implies a sorting algorithm, as well as reduction operations to compute the amount of data to be sent to each process. After that, the data is sent by the network to the rest of processes, which would take some time, depending on the amount of data to send, and the network speed.

In parallel, the tool will prepare everything to download the data incoming from the other processes, INSIDE THE SAME ARRAYS.

To reduce the computational cost and avoid aside effects, it is strongly recommended to copy the actual data into helper arrays before synchronizing.

Note: Since the mask array shall be sorted, power of 2 arrays are required.

Constructor & Destructor Documentation

◆ MPISync()

Aqua::CalcServer::MPISync::MPISync	(	const std::string	name,
		const std::string	mask,
		const std::vector< std::string >	fields,
		const std::vector< unsigned int >	procs,
		bool	once = `false`
	)

Constructor

Parameters

name	Tool name
mask	Mask of the data to be sent to each process. Numbers out of bounds (i.e. bigger or equal to the number of processes) will be ignored, and therefore not sent anywhere
fields	Fields to be synchronized between processes
procs	Processes to be considered to send information. If an empty list is provided, all the processes will be considered. Providing a list of processes is reducing the number of reductions to be carried out, and therefore the computational cost
once	Run this tool just once. Useful to make initializations

Warning: The used mask will be overwritten

◆ ~MPISync()

Aqua::CalcServer::MPISync::~MPISync ( )

Destructor.

Member Function Documentation

◆ _execute()

cl_event Aqua::CalcServer::MPISync::_execute ( const std::vector< cl_event > events )

protectedvirtual

Execute the tool

Parameters

events List of events that shall be waited before safe execution

Returns: OpenCL event to be waited before accessing the dependencies

Reimplemented from Aqua::CalcServer::Tool.

◆ setup()

void Aqua::CalcServer::MPISync::setup ( )

virtual

Initialize the tool.

Reimplemented from Aqua::CalcServer::Tool.

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

/home/pepe/SPH/Code/aquagpusph/include/CalcServer/MPISync.h

Data Structures

Public Member Functions

Protected Member Functions