|
EvolvingObjects
|
|
EvolvingObjects
|
Class implementing the centralized job algorithm. More...
#include <eoMpi.h>
Classes | |
| struct | FinallyBlock |
| Finally block of the main algorithm. More... | |
Public Member Functions | |
| Job (AssignmentAlgorithm &_algo, int _masterRank, int _workerStopCondition, JobStore< JobData > &_store) | |
| Main constructor for Job. | |
| void | run () |
| Launches the job algorithm, according to the role of the host (roles are deduced from the master rank indicated in the constructor). | |
| bool | isMaster () |
| Returns true if the current host is the master, false otherwise. | |
Protected Member Functions | |
| void | master () |
| Master part of the job. | |
| void | worker () |
| Worker part of the algorithm. | |
Protected Attributes | |
| AssignmentAlgorithm & | assignmentAlgo |
| int | masterRank |
| const int | workerStopCondition |
| bmpi::communicator & | comm |
| JobStore< JobData > & | store |
| SendTaskFunction< JobData > & | sendTask |
| HandleResponseFunction< JobData > & | handleResponse |
| ProcessTaskFunction< JobData > & | processTask |
| IsFinishedFunction< JobData > & | isFinished |
| bool | _isMaster |
Class implementing the centralized job algorithm.
This class handles all the job algorithm. With its store and its assignment (scheduling) algorithm, it executes the general algorithm described above, adding some networking, so as to make the global process work. It initializes all the functors with the data, then launches the main loop, indicating to workers when they will have to work and when they will finish, by sending them a termination message (integer that can be customized). As the algorithm is centralized, it is also mandatory to indicate what is the MPI rank of the master process, hence the workers will know from who they should receive their commands.
Any of the 3 master functors can launch exception, it will be catched and rethrown as a std::runtime_exception to the higher layers.
| eo::mpi::Job< JobData >::Job | ( | AssignmentAlgorithm & | _algo, |
| int | _masterRank, | ||
| int | _workerStopCondition, | ||
| JobStore< JobData > & | _store | ||
| ) | [inline] |
Main constructor for Job.
| _algo | The used assignment (scheduling) algorithm. It has to be initialized, with its maximum possible number of workers (some workers referenced in this algorithm shouldn't be busy). See AssignmentAlgorithm for more details. |
| _masterRank | The MPI rank of the master. |
| _workerStopCondition | Number of the message which will cause the workers to terminate. It could be one of the constants defined in eo::mpi::Commands, or any other integer. The user has to be sure that a message containing this integer will be sent to each worker on the Commands channel, otherwise deadlock will happen. Master sends Finish messages at the end of a simple job, but as a job can happen multiples times (multi job), workers don't have to really finish on these messages but on another message. This is here where you can configurate it. See also OneShotJob and MultiJob. |
| store | The JobStore containing functors and data for this job. |
| void eo::mpi::Job< JobData >::master | ( | ) | [inline, protected] |
Master part of the job.
Launches the parallelized job algorithm : while there is something to do (! IsFinished ), get a worker who will be the assignee ; if no worker is available, wait for a response, handle it and reask for an assignee. Then send the command and the task. Once there is no more to do (IsFinished), indicate to all available workers that they're free, wait for all the responses and send termination messages (see also FinallyBlock).
Definition at line 658 of file eoMpi.h.
Referenced by eo::mpi::Job< void >::run().
| void eo::mpi::Job< JobData >::worker | ( | ) | [inline, protected] |
Worker part of the algorithm.
The algorithm is more much simpler: wait for an order; if it's termination message, leave. Otherwise, prepare to work.
Definition at line 709 of file eoMpi.h.
Referenced by eo::mpi::Job< void >::run().
1.7.6.1 
