Performance & Reliability Oriented Combined File, Capacity Allocation on Distributed Systems

Abstract

Distributed Computing Systems (DCSs) have the potential for high reliability. When the topology of a DCS is fixed, the DCS reliability depends mainly on the allocation of various resources. One of the important resources to be allocated on a DCS are various files. Another problem of significance deals with the allocation of capacities to the links of a DCS such that the performance (throughput) is maximized and a cost constraint is satisfied. In this paper, we have developed a performance based reliability oriented file and capacity allocation scheme for distributed systems. In this scheme capacities are allocated to the links of the DCS and the files are allocated to the nodes of the DCS such that the overall effectiveness of executing a program which requires files from remote node(s) is maximized. A Genetic Algorithm (GA) based approach is used to solve this problem. The effectiveness of the GA based approach is demonstrated by comparing the results with the results obtained using exhaustive search of the problem state space. The paper also studies the impact of the Probability of Crossover (Pc) and the Probability of Mutation (Pm) on the results obtained using the GA. 1. The File and Capacity Allocation Problem The reliability of executing various applications (programs) in a Distributed Computing System (DCS) depends on the topology of the DCS, the program locations, and the allocation of files on the DCS [l]. The system reliability can also be improved by introducing redundancy in the system [2,3]. If the topology of a DCS is fixed then the overall reliability of executing various applications depends mainly on the file allocation on the DCS [2]. If there are n processing nodes and m data files in the DCS, then the total number of possible assignments are nm. Thus the optimal allocation of files on the processing nodes is a problem of exponential complexity [2]. The file allocation problem is formulated in terms of cost factors and constraints, the objective being to allocate files such that the cost factor@) are minimized/maximized and all the constraints are satisfied [2]/ The cost factor(s) are represented by a cost funciton. The nature of a cost function and the constraints is problem dependent. For example, system reliability or htroughoput are important in communication networks, while response time is extremely important in real time systems. *This research is in part supported by the President’s Grant at the University of Louisville. 0-7803-1814-5194 $4.00 @ 1994 IEEE Sanjay P. Ahuja Department of Mathematical Sciences SUNY at Oneonta Oneonta, NY 13820 Conventional measures such as TR [4], survivability index [5], and computer network reliability (CNR) [6] compute fault tolerance of a DCS from a network component failure point of view only. The effect of file distribution on the DCS reliability is ignored. The parameters Distri-buted Program Reliability (DPR) [7] and Distributed System Reliability (DSR) [7] take into account the effect of file distribution on DCS reliability and hence are more suited to evaluating the reliability of program execution. The problem of file allocation on a DCS such that the DPR/DSR is maximized has been solved in [2]. It is to be noted, however, that the DPR/DSR reliability parameters do not take into consideration the capacity of the communication links and the total installed system capacity. The traffic requirement of the system is also ignored. It is therefore implicitly assumed that all the links of the DCS are always capable of the required flow. This is an invalid asumption since the links of a DCS have a finite, preassigned capacity. The problem of channel (link) capacity allocation is another problem of significance and deals with allocating capacity to the different links of the DCS whose topology is already fixed. The objective of this problem is to allocate capacity to the links of the DCS such that the throughput is maximized and the constraints are also met. One possible constraint could be that the total cost of installing capacities to the various links of the DCS not exceed a specified upper bound. Both these problems, i.e. file allocation with reliability optimization and capacity allocation have been solved individually in the literature [2]. However, while allocating files on a DCS such that DPR/DSR is maximized, the capacity allocation of the links is ignored, whereas the link capacities have a significant impact on the overall effectiveness of the DCS. On the other hand while solving the capacity assignment problem such that the throughput is maximized, the effect of file allocation on the throughput is ignored. Such single criterion optimization does not yield a DCS with a truly optimum file or capacity assignment. It is clear that these two criteria need to be optimized together to yield a truly optimal file and capacity assignment such that the overall effectiveness of the DCS is maximized. One way to accomplish this would be to perform a multi-criteria optimization. However, the newly developed parameter, Average Distributed Program Throughput [8] yields itself very well in solving this problem as it takes into