Performance Prediction Model for Distributed Applications on Multicore Clusters
Nontokozo Portia Khanyile, Jules‐Raymond Tapamo, Erick Dube · 2012
Distributed processing offers a way of successfully dealing with computationally demanding applications such as scientific problems. Over the years, researchers have investi- gated ways to predict the performance of parallel algorithms. Amdahl's law states that the speedup of any parallel program has an upper bound which is determined by the amount of time spent on the sequential fraction of the program, no matter how small and regardless of the number of processing nodes used. This research discusses some of the short comings of this law in the current age. We propose a theoretical model for predicting the behavior of a distributed algorithm given the network restrictions of the cluster used. The paper focuses on the impact of latency and bandwidth which affect the cost of interprocessor communication and the number of processing nodes used to predict the performance. The model shows good accuracy in comparison to Amdahl's law. ISTRIBUTED systems are collections of autonomous computing systems which are connected by some net- work and work together to solve an overall task using the notion of divide and conquer. The total processing time of a distributed program is calculated as the total communi- cation time plus the total computation time. It is always unnerving to know what to expect from a system before it is actually developed. Clients often need to know the expected performance so that they can make an informed decision on whether or not to invest in a project. Amdahl came up with a law for predicting performance of parallel systems. However, there has been much criticism surrounding Amdahl's law for its assertion that parallel processing is unscalable. Amdahl's law stipulates that even when the serial fraction of a problem, say s, is considerably small, the maximum speedup obtain- able is only 1 even with an infinite number of processing nodes (1). If s is the time spent by N processors executing the serial fraction of the computation time of a program and p is the time spent executing the parallel portion, then Amdahl's law states that the estimated speedup is given by: Speedup = 1 (s + p N ) ; with s=1-p (1)