Performance Prediction of Parallel Low-level Image Processing Operations
|
Abstract:
Abstract. An analytical performance prediction method is presented for predicting the performance of transputer based low-level image processing operations. The method is based on the fixed-sized speedup formula and uses a system overhead function to represent the communication and computation overhead as well as the hardware and implementation dependent (e.g. topology) characteristics of the parallel system. To determine the effect of communication on performance, first the properties of the transputer link are investigated to create a simple link model, then the communication patterns of the data-parallel algorithms are analysed to derive the overhead function. Using this overhead function, the parallel execution time, speedup and scalability of a given algorithm can be predicted in advance, without system building, helping developers to design efficient and scalable transputer-based parallel programs.
Citations
| 62 | Isoefficiency: Measuring the scalability of parallel algorithms and architectures – Grama, Gupta, et al. - 1993 |
| 43 | Occam 2 Reference manual – Ltd - 1988 |
| 41 | Analytical performance prediction on multicomputers – Clement, Quinn - 1993 |
| 26 | Multiprocessor Performance – Gelenbe - 1989 |
| 19 | The Search for Lost Cycles: a New Approach to Parallel Program Performance Evaluation – Crovella, LeBlanc - 1993 |
| 2 | The Transputer Databook – Ltd - 1989 |
| 1 | Measuring and Analyzing Parallel Computing Scalability – Zhang, Yan, et al. - 1994 |
| 1 | Zemerly et al., Characterising Computational Kernels to Predict Performance of Parallel Systems, in A. De Gloria et al – J - 1994 |
| 1 | Efficient Communication Methods for Minimising Overhead in Parallel Image Processing Algorithms – Juhasz - 1994 |
