:
:
:
: 2016
: ..
:
: 64
: . . // . 64. .: , 2016. . 81-111.
:
: , , , ,
(.): hypercubic computing system, supercomputer, parallelism, topological scalability, edge scaling
: : , , . , . , .
(.): The problem we consider is how to increase a parallelization potential of tasks in a computing system not extending the number of its processor units. Our approach is to introduce additional links to initial system topology but preserve all existing links, addressing and routing policies. We provide a general formulation of the proposed approach and detailed description for systems with hypercube topology. In an edge scaled hypercube the conventional addressing scheme is enhanced by inverse addressing. We demonstrate that bisection bandwidth of the modified network is twice as high as of conventional hypercube and the diameter is twice as low. We obtain the scalability functions of a limit parallelization and a topological scalability on tasks with limited path length between branches. The proposed network modification can be applied to a limited k-dimensional fragment of the network with adaptive k selection. The impact on reliability of such systems will be covered in future research.
PDF -
: 3463, : 1284, : 18.