Toon, thank you! I will give it a try here so we can have some data points.
Jerry On Thu, Jun 26, 2025, 2:08 PM Toon Moene <t...@moene.org> wrote: > On 6/26/25 21:34, Thomas Koenig wrote: > > > Am 26.06.25 um 10:15 schrieb Andre Vehreschild: > > >> Hi Thomas, > >> > >>> I have a few questions. > >>> > >>> First, I see that your patch series does not use gfortran's descriptors > >>> for accessing coarrays via shared memory, as the original work by > >>> Nicolas did. Can you comment on that? > >> > >> The ABI for invoking coarray functionality is sufficient for doing the > >> job. > >> Modifying the compiler to access coarrays directly, i.e., having > >> implementation > >> detail on a certain library in the compiler did not appeal to me. > >> Furthermore > >> has the new library in conjunction with the other library available the > >> potential to get to a stable and maintained ABI. Having another ABI in > >> the > >> compiler would have lead to two badly maintained ones (in my opinion). I > >> therefore decided to just have one ABI and figured that all that is > >> needed can > >> be done in a library. This also allows link-time polymorphism. And > >> last but not > >> least, there is a major rework on the array descriptor going on and > >> that would > >> have had the potential to conflict with my work. > > > > We are very probably not going to get performance out of it that is > > comparable with the original design; I would be quite surprised if it > > was appreciably better than using shared memory MPI, and in that > > case I don't see an advantage of your patch over a better wrapper. > > When I learned about the "Koenig" implementation of "native coarrays" in > 2018 (as they called it at that time, and I since noticed also Intel > calls it), I wrote a "mock weather forecasting program" using coarrays > to test it against the then working implementation of OpenCoarrays using > MPI calls. > > You can find the program here: https://moene.org/~toon/random-weather > > [ note that I improved on this program until early in 2021. ] > > When I compared the run time of the two implementations with the same > input parameters on a 128 Gbyte RAM Intel machine, the "native" > implementation was around a factor of 5 faster. Of course, the > OpenCoarrays based MPI implementation (using OpenMPI) used shared memory > MPI (which OpenMPI calls "vader" for reasons that escape me). > > So I am certainly interested to compare Andre's implementation against > OpenCoarrays. > > Kind regards, > > -- > Toon Moene - e-mail: t...@moene.org - phone: +31 346 214290 > Saturnushof 14, 3738 XG Maartensdijk, The Netherlands >
