The Standard Performance Evaluation Corp. is a non profit, industrial consortium that maintains many kinds of system benchmarks. The Server File System benchmark, SPEC SFS97 , measures file server performance. The SFS97 figure of merit is the change in response time vs. increasing load on the server. Systems can thus be compared based on response time, throughput, as well as scalability. SPEC)
This project would observe the effects of changing the definition of "load" for the SFS97 benchmark. Currently, the SFS clients use a poisson arrival rate to model load on the server. However, a recent paper has shown that for the timescales on the order of the length of the benchmark (minutes), file operation arrival rates are better characterized by self-similar arrival rates. One method of generating self similar traffic is discussed in the above self-similar paper, while another self-similar traffic method for web servers is described in this paper.
Alter the SFS clients to generate a workload on a server. Such workloads can be approximated by using the sum of heavy tailed distributions. How does the average response time vs. load curve change in respect to the different classes of workload? More formally, how could you characterize the distribution of response time for a given load. For example, even if a server had a good average response time, a high variance may be unacceptable. Is the maximum server throughput under a poisson arrival close to what the server can sustain under a self-similar workloads?
How representative are the file size access patterns used by SFS97? Can you find data to describe how similar or dissimilar the current SFS97 file access model is? One paper that describes file access patterns can be found here.