Gulley, Jonathan William; (1998) Methods for Improving the Performance of Software for DM MIMD Systems. Doctoral thesis (Ph.D), UCL (University College London).

Text Methods_for_improving_the_perf.pdf Download (7MB)

Abstract

In high-performance computing, distributed-memory multiple-instruction multiple-data systems, in which processors have local memories and exchange data via links, have the potential to achieve very high degrees of parallelism at low levels of cost. The use of non-deterministic, complex software behaviours, in order to guard against performance risks, has raised the cost of realising this potential. Programs with simpler behaviours are much easier to construct, debug and modify, but their success depends upon being able to manage performance risk successfully, particularly through tuning after implementation. At the current state of the art, program tuning is supported only by performance monitoring tools, which give no prediction of the performance which will be achieved by tuning. The performance risks in program design, on the other hand, are reduced by the analysis and simulation of performance models. This dissertation extends the field of research by evaluating methods which bring model analysis and simulation to bear on the problem of program tuning. For some programs, model analysis can predict, on the basis of monitoring results, limits on what can be achieved through tuning for various levels of cost. Model simulations help to determine whether these gains will, in fact, be realized. For an idealised parallel program, the benefit of this technique is that the costly life cycle iterations associated with tuning parallel programs are greatly reduced, because performance is predicted before the high costs of tuning are incurred. The obstacles of program complexity, data-dependent grain sizes, parallel program structure and imperfect performance tools impose limitations on the technique, though these limitations are partially addressed by proposing novel extensions to the tools and techniques. Further research is needed to determine to what extent realistic applications can be developed within the class of deterministic programs which can be addressed successfully.

Download activity - last month

Export as CSVJSONXML

Download activity - last 12 months

Export as XMLJSONCSV

Downloads by country - last 12 months

Export as CSVXMLJSON

Archive Staff Only

View Item