Self-Tuning Systems

Modern operating systems are highly parameterized, allowing system administrators to tune them in order to achieve optimal performance for the local workload. However, this is a diicult and time consuming process. We propose a mechanism to automate this process, by running simulations of system performance for various parameter values instead of the system's idle loop. The simulations are driven by log les containing information about the local workload, and genetic algorithms are used to search for the optimal parameter values. We evaluated this idea by running such simulations oo-line. A case study involving batch scheduling on an iPSC hypercube found parameter values that reduced fragmentation and salvaged a quarter of the computing cycles that were lost when using the default values.