types of parallel computers, clusters

We started the lecture with an example of deadlock when messages are sent in packets through a network. The e-cube routing algorithm (derived from the hypercube architecture) breaks the cycles and thus prevents deadlock. Via switches, dynamic networks are realized. We derived an omega network for eight processors.

Flyn's SISD and MIMD classification of parallel computers was defined.

A cluster is an independent set of computers combined into a unified system through software and networking. Beowulf clusters are scalable performance clusters based on commodity hardware, on a private network, with open source software infrastructure. The driving forces behind the clustering revolution are commodity hardware, the Ethernet networking, and open source software: Linux, MPI, LApack.

Notice that -- while a cluster computer can achieve a performance worthy of a supercomputer -- a cluster computer is not a parallel computer. The key words marking the difference are "commodity" and "dedicated". A parallel computer may have specialized hardware, a specialized interconnection network (Myrinet instead of Ethernet), and specialized operating systems and system software, dedicated to achieve an optimal performance.

Bibliography