GASPI is focused on an asynchronous data-flow model for distributed memory architectures. Whereas in MPI one-sided communication happens in specific time intervals (called an epoch), GASPI does not require such a concept. Rather, GASPI makes use of remote completion in the form of notifications. Data may be written asynchronously whenever it is produced, accompanied by a corresponding notification, and data is processed locally whenever the required preconditions for processing this data can be met, i.e. whenever the corresponding notifications have been flagged. The GASPI notification mechanism hence allows for a highly scalable asynchronous data-flow model, which is not only able to provide a very fine-grained overlap of communication and computation, but is also very robust w.r.t. system jitter and noise.
Why should GASPI scale better than MPI? Isn’t one-sided communication in MPI-3.0 identical to the GASPI communication?
