Bulk-synchronous communication does not scale. An application-specific asynchronous gather/scatter collective which overlaps communication with computation in an application-specific way is not only easy to implement in GASPI but in most cases would outperform the generic implementation that GASPI might offer. Since we also want to keep the GASPI interface as small as possible we have deliberately removed these collectives from the GASPI API.
The set of collectives in GASPI appears to be somewhat limited – why is that?
