City Research Online

Abstract

Increased computer throughput can be achieved through the use of parallel processing. The granularity of a parallel program is the average number of instructions performed by the tasks constituting it. Coarse-grained programs typically execute huge numbers of instructions per task (w 105). The tasks in fine-grained programs are typically short (æ 103). In general, the finer the program grain, the greater the potential for exploiting parallelism. Amdahl’s Law shows that in the absence of overheads, the more potential parallelism that is realised in an algorithm, the faster it will be. The economical granularity of tasks is determined by the intertask communications overhead. Break-even occurs when processing is approximately equally divided between useful work and overhead.

The two common parallel programming paradigms are shared variable and message passing. Shared variable is, in general, the more natural of the two as it allows implicit communication between tasks. This encourages the programmer to make use of fine-grained tasks. The message passing paradigm requires explicit communication between tasks. This encourages the programmer to use coarser-grained tasks.

Two kinds of parallel architecture have become established. The first is the multiprocessor, which is built around a shared bus giving broadcast communications and a shared memory. This is characterised by low communications overhead, but limited scalability. The second is the multicomputer, which is based on point-to-point communications with larger communications overhead, but good scalability. Quantitatively, the low overhead of the multiprocessor is well matched to fine-grain tasks and, hence, to supporting the shared variable paradigm, while the high overhead of the multicomputer matches it to coarse-grain parallelism and, hence, to the message passing paradigm.

Currently, there appears to be no middle ground in parallel computing; an architecture which can support both several hundred medium-grained (« 104 instructions) parallel tasks and the shared variable programming paradigm would be advantageous in many applications.

This thesis asserts that it is possible to implement a new computer architecture, Cherub, which has at least 200 processors and is able to support shared variable programming with an optimal task granularity of around 104 instructions. This can be achieved through the combination of a hardware-based distributed shared single address space and a wafer-scale communications network.

To support the thesis, the dissertation first specifies a programmer’s interface to Cherub which is simple enough to implement in hardware. It then designs algorithms which provide this interface, allowing the requirements of the underlying network to be estimated. Finally, a wafer scale communications network is outlined, and simulations are used to demonstrate that it can provide the performance required to successfully implement Cherub.

Publication Type:	Thesis (Doctoral)
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Departments:	School of Science & Technology > Computer Science School of Science & Technology > School of Science & Technology Doctoral Theses Doctoral Theses

Export

Downloads