The segments in GASPI are closely related to the MPI windows used in one-sided MPI communication. However, in GASPI these segments can be freely configured. This enables GASPI users to map the memory heterogeneity of a modern supercomputer node to GASPI segments. As examples, GASPI allows users to map the main memory of a GPGPU or Xeon Phi to a specific segment, to configure a GASPI segment per memory controller in a CC-NUMA system or to map non-volatile RAM to a specific segment. All these segments can directly read and write from/to each other – within the node and across all nodes.
What are segments in GASPI? How can I use them? How do they differ from MPI windows?
