Resource Allocation for Real-Time Tasks in High Performance Distributed Computing Systems

High performance distributed computing paradigm offers various types of allocation mechanisms to resource-intensive user tasks. To achieve a high level of confidence in temporal constraints and high throughput, scheduling mechanism at task level is of vi tal importance in resource allocation. The task scheduling problem has become more complex with the nature of tasks and the ever increasing size of high performance computing systems. Even though selecting an efficient resource allocation strategy for a particular task helps in obtaining a desired level of service, researchers still face difficulties in choosing a suitable technique from a plethora of existing methodologies in literature. In the present dissertation, we explore and discuss existing resource al location mechanisms for resource allocation problems employed in high performance distributed systems. The work comprehensively surveys and compares resource al location mechanisms for different architectures (centralized, distributed, static or dynamic) based on different parameters such as time complexity, searching mecha nism, allocation strategy, optimality, operational environment and objective function they adopt for solving computing- and data-intensive applications. Resource alloca tion mechanism in high performance distributed computing systems becomes more challenging when computationally intensive tasks have real-time deadlines. Such re source allocation mechanism maps tasks to the available resources according to some predefined criteria such as minimizing makespan or execution cost, load balancing, energy efficiency, maintaining user defined tasks deadlines, efficiently using resource memory etc. Makespan minimization is a dominant criterion which becomes more complex when real-time tasks have data requirements. To obtain feasible results, such tasks need data files to be processed within deadlines. The required data files are transferred from data storage resources to the computing resources which con sume network bandwidth. Resource allocation mechanism for such tasks takes into account the data files transfer time and processing power of computing resources to x complete execution within deadlines. This dissertation explores the problem of allo cating real-time data-intensive tasks to Grid heterogeneous computing resources with the assumption that data resources are decoupled from the computing resources and connected by network links of various bandwidths. The resources are analytically formulated with the aim to simultaneously maximize total number of mapped tasks in a metatask that guarantees execution of the tasks within deadline constraints with minimum possible makespan. Similarly, the applicability of Cloud computing services for real-time systems, especially for hard real-time systems where deadlines must be observed under all circumstances and providing adequate resource mapping criteria is still required. In current dissertation, we also propose a platform for real-time systems in Cloud environment by addressing the scheduling problem. The proposed mechanism acquires Virtual Machines for a specific period of time to satisfy all task constraints and increase the utilization of available background resources. The cost effective task scheduling and Virtual Machines allocation problem for real-time tasks is also solved by using hybrid heuristic approach. The proposed mechanisms are validated via simulations and mathematical formulations. The results show that all tasks meet their deadlines in terms of selected metrics when run in high performance computing systems with our proposed solution.