It would be nice if a single dspsr process could be run on multiple CPUs/cores using multiple threads. These threads could share important resources, such as the FFT plans, dispersion kernel, and phase predictor, without the need for MPI. Potential trouble spots:
- a mutex would be required to protect shared resources during allocation/initialization;
- the InputBuffering system would have to implement a shared pool, such that each thread could store the data that it lost for use by other threads; and
- static variables, such as the Operation::workingspace, would have to be removed.
Logged In: YES
user_id=1219749
Originator: YES
Multi-threaded dspsr is pretty much working now and any problems should be registered as bugs.