Abstract:
Architectural aspects of exascale supercomputers are explored. Parameters of the computing environment and interconnect are evaluated. It is shown
that reaching exascale performances requires hybrid systems. Processor elements of
such systems comprise CPU cores and arithmetic accelerators, implementing the
MIMD and SIMD computing disciplines, respectively. (in Russian)
Efficient exascale hybrid systems require fundamentally new applications and
architectural efficiency scaling solutions, including:
process-aware structural reconfiguring of hybrid processor elements by
varying the number of MIMD cores and SIMD cores communicating with
them to attain as high performance and efficiency as possible under given
conditions;
application of conflict-free sets of sources and receivers and/or decomposition of the computation to subprocesses and their allocation to environment
elements in accordance with their features and communication topology to
minimize communication time;
application of topological redundancy methods to preserve the topology and
overall performance achieved by the above communication time minimization solutions in case of element failure thus maintaining the efficiency
reached by the above reconfiguring and communication minimization solutions, i.e. to provide fault-tolerant efficiency scaling.
Application of these solutions is illustrated by running molecular dynamics
tests and the NPB LU benchmark.
The resulting architecture displays dynamic adaptability to program features,
which in turn ensures the efficiency of using exascale supercomputers.
Key words and phrases:Hybrid architectures, architectural efficiency scaling solutions, hybrid reconfigurable structures, minimization of communication time, topological redundancy.