Parallel I/O subsystems for distributed-memory multicomputers

The authors study the improvement in performance obtained in distributed memory machines through the use of a separate network that serves multiple I/O nodes operating under a distributed file system. For a hypercube architecture augmented by an independent I/O network that scales as N/log N, a significant improvement in performance is observed, specially when data locality is low. Moreover, performance becomes relatively insensitive to data locality. From the hardware aspect, wormhole routing is efficient only for low network loads, and its performance degrades rapidly even with moderate channel utilisation. Simulation results are presented for a 128-node hypercube