Running an MPI application requires using the mpirun command. The mpirun command allows you to, among other things, specify the number of processes that should be started in the application. mpirun then sets up the MPI communication infrastructure and starts the processes in a way that the MPI API knows how to communicate.
If you had to set the OPAL_PREFIX environment variable, you can use it. Here's a sample invocation of the mpirun command that starts 5 processes for the SpecTcl that you built in the current working directory:
the value given to the -n option specifies the number of processes mpirun
should start. As will be described more completely in
How mpiSpecTcl works in parallel mode., you must use at least 3 processes.
Two processes are non-worker processes, the remaining processes are workers that run the event processing
pipeline in parallel on events. The number actual value of -n that makes sense,
depdends on the computational complexity of your event processing pipeline.
If you use this simple invocation of the miprun command your processes will all run
in the computer that runs the mpirun command. In that case,
a value for -n larger than the number of cores in that system
makes no sense. More complex invocations of mpirun are possible and
can allow the application to spread across more than one system. This is complicated for
the containerized environment and beyond the scope of this document.
Events read from the data source are dispatched to the worker processes. The number of
worker processes are 2-n where n is the value given to the -n
option of mpirun.
Events are dispatched to workers in chunks. The number of ring items (events) in each chunk can be controlled by setting the WorkerChunkSize Tcl variable to the desired number of ring items per chunk. While this is normally set in SpecTclInit.tcl, after chunks are sent, SpecTcl dynamically updates the chunksize fromt he current value of that variable. If this variable is not set or is not a legal integer greater than 0, it defaults to 1