The serial mode was for some time now probably seen as a "legacy" part since it referred to the first implementation that we put before the parallel mode.
I would propose, in case we want to refactor that in the future, that the serial mode could be evolved to something similar to the aggregated mode in ADIOS, where we use several parallel writes but in more than one file. But we would not use explicit aggreagators for that but directly MPI parallel writes in MPI sub-communicators.
This would be very neat, since we could re-use all our "find and merge" algorithms and at the same time overcome some of the limits of MPI-I/O that are coming from writes to single files. The non-MPI version of libSplash would in that case be dropped and both versions would use MPI, but the multi-file-MPI version would be better in performance. Comments are welcome.
The serial mode was for some time now probably seen as a "legacy" part since it referred to the first implementation that we put before the parallel mode.
I would propose, in case we want to refactor that in the future, that the serial mode could be evolved to something similar to the
aggregatedmode in ADIOS, where we use several parallel writes but in more than one file. But we would not use explicit aggreagators for that but directly MPI parallel writes in MPI sub-communicators.This would be very neat, since we could re-use all our "find and merge" algorithms and at the same time overcome some of the limits of MPI-I/O that are coming from writes to single files. The non-MPI version of libSplash would in that case be dropped and both versions would use MPI, but the multi-file-MPI version would be better in performance. Comments are welcome.