A multi-level direct interconnect
topology, also known as dragonfly, might be a viable option for exascale
supercomputers. However, it raises interesting questions about job scheduling,
routing and task mapping.
Step-by-step embedding of one
two-dimensional catesian space into another (using algorithm discussed in Rosenberg et
A finer domain decomposition for molecular
dynamics leads to communication with more and more neighboring processes.
However, at exascale, all of them might be placed on cores of the same node
thereby not increasing the intra-node communication requirements.