Optimization of Workload Allocation Problem in a Network of Heterogeneous Computer Systems

Multiple Queue Multiple Server Queueing models are used to model workload allocation problems in a network of computers. The problem of determining optimal allocation of workload with single and multi class jobs to a parallel of computers is presented followed by a problem of determining optimal buffer size related to arrival of
workload to a single computer. The generalized exponential (GE) distributional model with known first two moments has been used to represent general inter arrival and service time distributions as various jobs have various traffic characteristic. Other service distributional models such as exponential, Erlang-k and Gamma have also been used to expand the work applicability. A new algorithm of workload allocation scheme using First Come First Serve discipline in conjunction with optimization of GE queueing systems is proposed for minimizing mean queue length and mean response time in a network of computer systems. This has an advantage over a classical queueing allocation scheme, and is favorably compared. The performance measures, mean queue length and mean response time of the proposed scheme have practically shown improvement.
The principle of optimization and GE distributional model are used to derive a new workload allocation model of single class jobs in a network of queueing system. The
reusable concept is proposed to gain solution for determining individual job allocation in a multi class environment. This study is of interest whereby both single and multi class assumption can be done without repeatedly developing and solving new models.
The convincing results of workload allocation models proposed has motivated the work to obtain the direct dependence of the buffer size on the given workload in the network of computer systems. A closed loop expression for buffer sizing of single class jobs and partial buffer sizing of multi class jobs are derived and show their dependency on workload arrival and processing rate in a computationally efficient way.