tarch::mpi Namespace Reference

Namespaces
namespace	tests

Data Structures
class	BooleanSemaphore
	Boolean semaphore across MPI ranks. More...
struct	DoubleMessage
struct	IntegerMessage
class	Lock
	Create a lock around a boolean semaphore region. More...
class	Rank
	Represents a program instance within a cluster. More...
struct	StringMessage
	The string message looks like the other messages which I generate through DaStGen 2, but I actually wrote it myself manually to support dynamic string lengths. More...

Functions
std::string	MPIStatusToString (const MPI_Status &status)
	Returns a string representation of the mpi status.
std::string	MPIReturnValueToString (int result)
void	wait (MPI_Request &request, const std::string &fullQualifiedClassName, const std::string &functionName)
	Wrapper around MPI_Wait.

Detailed Description

Using DaStGen2's send and receive

With DaStGen2, we have two variants of sends and receives: We can use the blocking variant and a non-blocking one. Furthermore, we can obviously use the MPI datatype that the generator yields as well. With the blocking variant, the usage is straightforward:

IntegerMessage::receive(message, rank, BarrierTag, tarch::mpi::Rank::getInstance().getCommunicator());
 

With the option to insert a wait functor, you can do more complicated stuff. The minimalist version is:

tarch::mpi::IntegerMessage::receive(
  message, rank, BarrierTag,
  []() {},
  [&]() {
    tarch::services::ServiceRepository::getInstance().receiveDanglingMessages();
  },
  tarch::mpi::Rank::getInstance().getCommunicator()
);
 

It ensures that MPI progress is made as it actively polls the MPI subsystem for further incoming messages. I do prefer the more sophisticated version which also has a timeout mechanism:

 tarch::mpi::IntegerMessage::receive(
   message, rank, BarrierTag,
   [&]() {
     tarch::mpi::Rank::getInstance().setDeadlockWarningTimeStamp();
     tarch::mpi::Rank::getInstance().setDeadlockTimeOutTimeStamp();
   },
   [&]() {
     tarch::mpi::Rank::getInstance().writeTimeOutWarning( "tarch::mpi::DoubleMessage", "sendAndPollDanglingMessages()",destination, tag );
     tarch::mpi::Rank::getInstance().triggerDeadlockTimeOut( "tarch::mpi::DoubleMessage", "sendAndPollDanglingMessages()", destination, tag );
     tarch::services::ServiceRepository::getInstance().receiveDanglingMessages();
   },
   tarch::mpi::Rank::getInstance().getCommunicator()
 );

As inserting the code from above is cumbersome, I decided to create my own Peano4 aspect and to inject it into DaStGen. So just invoke the sendAndPoll or receiveAndPoll operation and you basically get the above aspect.

Function Documentation

MPIReturnValueToString()

std::string tarch::mpi::MPIReturnValueToString

(

int

result

)

Referenced by peano4::stacks::STDVectorStack< T >::startReceive(), and peano4::stacks::STDVectorStack< T >::startSend().

Here is the caller graph for this function:

MPIStatusToString()

std::string tarch::mpi::MPIStatusToString

(

const MPI_Status &

status

)

Returns a string representation of the mpi status.

For a detailed description of the mpi status itself see the file mpi.h.

wait()

void tarch::mpi::wait	(	MPI_Request &	request,
		const std::string &	fullQualifiedClassName,
		const std::string &	functionName )

Wrapper around MPI_Wait.

This routine wraps around MPI_Wait and implements some deadlock detection. That's the not the only thing: the routine also invokes the receiveDanglingMessages() routine.

Parameters

request	Valid MPI request
fullQualifiedClassName	The fully qualified function name invoking the wait(). This one is used to report on deadlocks.