Analysis of resource scheduling strategies in parallel, distributed and grid computing systems

Large-scale parallel and grid computing systems are becoming more mainstream in order to achieve high performance. The handling of large volumes of computational load on distributed and parallel computing systems is becoming a more and more challenging task. In this dissertation, we consider the problem of the analysis of resource scheduling strategies in parallel, distributed and grid computing systems. We employ a most promising algebraic means of determining the optimal allocations of load fractions to processors and links in a given interconnection topology, Divisible Load Theory (DLT). We investigate scheduling on a real-time networks with arbitrary processor release times. The theoretical analysis shows that generated solutions based on a conjectured timing model does not always admit a feasible solution. A heuristic algorithm for generating sufficient models for a feasible solution is developed. For the next topic, scheduling scenarios in wireless sensor networks, which are dynamically growing and promising distributed computing systems, are examined based on DLT. The performance of these scenarios is examined with respect to different sensing speeds, communication speeds, and information utility constant parameters. The special bounds for the ratio of speed parameters for the maintenance of the minimum round time for certain scenarios are derived. Mathematical analysis based on DLT contributes to interpret and evaluate performance of the wireless sensor networks. For the next study, we consider a simulation of the adaptive capacity utilization problem for data transfers between multiple source and destination nodes interconnected by modern, high-performance, hybrid networks that support resource reservations. We account for a file transfer scenario using node capacity, file size, file transfer start time, and the deadline of files when determining the explicit capacity of Virtual Paths (VPs) across a backbone. Two opposite heuristic algorithms are designed to react to file transfers according to the temporal capacity of VPs for multiple sources and destinations networks. For the next topic, a special type of parallel computing system, "Grid" computing systems, has appeared as a promising trend for large-scale distributed parallel processing systems. An optimal computing power allocation solution is adapted to divisible load in a parallel computing grid is with an idealized two sources and a single sink computing processor. For the last topic, one of the major challenges to all processor requirements for high performance network systems now and in the future will be low cost. By applying DLT, we consider the problem of monetary network cost on a homogeneous tree networks. Our objective is to analyze quantitative and qualitative trends of monetary network cost against ratio of network speed parameters and to determine relationships between the network cost and the network environment. As going to Terascale/Petascale high performance computing (HPC) systems and beyond means that the number of components (cores, interconnect, storage) within such a system will grow enormously, it is obvious that these highly parallel systems will raise questions about reliable information about resource management and scheduling. Thus, it can be expected that adaptation of DLT is promising technique for a new era of HPC especially for resource allocation, computing power adjusted according to computing load, and network performance prediction in a parallel paradigm.