AN APPROXIMATE METHOD FOR DETERMINATION OF HIGH LEVEL LANGUAGE STATEMENT EXECUTION TIMES WITHOUT RECOURSE TO ASSEMBLER OR MACHINE CODE

• Published in: UKSC 84 pp. 118 - 131
• Main Author: BENNETT, B S
• Format: Book Chapter
• Language: English
• Published: Elsevier Ltd 1984
• ISBN: 9780408015042; 1483144577; 9781483144573; 1483112152; 9781483112152; 0408015047
• DOI: 10.1016/B978-0-408-01504-2.50018-1

During the course of digital computer simulation of continuous systems using a distributed network of processors, it is important that all the processors should be utilised as fully as possible to derive maximum benefit from multiple processing. The work-sharing among processors for such a simulation task is usually carried out by one of two methods:(a)Integration algorithm partitioning which requires special parallel integration algorithms capable of being computed simultaneously on all processors in the network.(b)Equation set partitioning which means that identical copies of the integration algorithm are implemented on all processors. Each processor then solves a sub-set of the model equations and communicates the results to the others in the network. In order to maximise the utilisation of all processors, it is essential to try to equalise the computation load between them as far as possible. This requires assessment of the time taken for derivative function evaluation, which, along with the integration algorithm computation, constitutes the greatest part of the total computational load. A method is put forward in this paper for the estimation of statement execution times without recourse to the counting of assembler or machine code instructions. effects of such coefficient differences. For the two models concerned, use of each other's sets of regression coefficients resulted in execution time estimates that did not match the actual times so well, but the differences were considered to be less important than the truncation effects arising from model equation set partition across whole numbers of equations. The work of this paper has been wholly confined to the six arithmetic operations. No attention has been paid to procedure or library function calls. These are likely to be highly unpredictable, and much work would need to be done to quantify their effects in execution timing assessment by this method.