A fair workload allocation policy for heterogeneous systems

  • Authors:

  • Leonidas Georgiadis;Christos Nikolaou;Alexander Thomasian

  • Affiliations:

  • Aristotle University, Department of Electrical Engineering, Faculty of Engineering, Thessaloniki, Greece;Computer Science Department, University of Crete, Heraklion, Crete, Greece;Computer Science Department, New Jersey, Institute of Technology, Newark, NJ

  • Venue:
  • Journal of Parallel and Distributed Computing
  • Year:
  • 2004

Quantified Score

Hi-index 0.00

Visualization

Abstract

We consider a new workload allocation policy addressing fairness for user level performance measures. More specifically the criterion used for optimal workload allocation is the one which minimizes the maximum expected response time at computer systems to which jobs are routed. The policy to attain this criterion is therefore referred to as the min-max policy (MMP). It is shown that this optimization criterion is tantamount to routing to the fastest M processors, where M depends on system statistics and equalizing the expected response times on these processors. The algorithm to compute job routing probabilities is applicable to increasing continuous functions of system response time versus the job arrival rate. We next investigate some properties of the MMP and show that it results in minimizing the coefficient of variation of response time when the job processing times are exponentially distributed. We compare the MMP with the one that minimizes the mean overall response time. It is shown that the new policy attains fairness by equalizing the mean response times at different systems, at a tolerable increase in overall response time. Finally, we report on a sensitivity analysis with respect to changes in job arrival rate and errors in estimating this rate.