Seminars & Colloquia
College of William and Mary
"Performance Impacts of Autocorrelation in Systems"
Monday December 04, 2006 10:30 AM
Location: 3211, EB II NCSU Centennial Campus
(Visitor parking instructions)
For the case of finite buffer systems, we use measurements from a 3-tier e-commerce server under the TPC-W workload and show the presence and propagation of autocorrelated flows in all tiers of the system, despite the fact that the stochastic processes used to generate this session-based workload are independent. We attribute this effect to the existence of autocorrelation in the service process of one of the tiers. In contrast to systems with independent flows, autocorrelation in the service process may result in very high user system response times despite the fact that bottleneck resources are not highly utilized, and measured throughput and device utilization levels are modest. This, falsely indicates that the system can sustain higher capacities. We present a small queuing network that helps us understand the above counter-intuitive behavior.
For a system with infinite buffer size, we consider the problem of load balancing in homogeneous cluster environments admitting jobs with high variability in their execution times. Load balancing in such environment has been shown to heavily depend on the policy's ability to direct jobs to servers according to the job size. The effectiveness of such policies is based on separating 'short' from 'long' jobs, by avoiding having short jobs waiting behind long jobs for service. We show that performance improvements due to this separation quickly vanishes if there the job arrival process to the cluster is autocorrelated. Based on this fact, we devise a new size-based policy (i.e., jobs are still directed to different servers according to their size) but all servers are not equally utilized so that the performance loss due to autocorrelated arrival flows is minimized.
Host: Harry Perros, Computer Science, NCSU