Applying Genetic Algorithms to Extract Workload Classes

It is often desirable to predict how a computer system will perform given changes to the system. Systems administrators are frequently faced with answering questions such as, \How will the throughput of the system be aaected if the number of users increases by 50% and if the CPU is upgraded by adding a oating point processing unit?" System models (e.g., analytical, simulation) can be used to help answer such questions. As one example, a closed multi-class queueing network model may be constructed of the system. This model consists of a network of servers and a set of customer classes. Each customer class is deened by the number of customers (i.e., jobs) in the class and by the demands that each customer in the class places on each of the servers. If the underlying model and assumptions are accurate, it is possible to predict how changes to the system will aaect performance. One typical assumption of queueing models is that the time spent at a server by a class is exponentially distributed. A primary workload characterization problem of such models is the determination of the customer demands that each customer class places on each device. This is a diicult measurement task. Software/hardware monitors can measure overall activity at a device, but it is diicult to associate a particular activity to a particular customer class. For instance, a monitor may be able to obtain the composite distribution of all customer class demands at each device, but not the distribution of each individual customer class. The classical method of extracting the individual customer class demands from the composite monitor data is clustering. However, typical clustering algorithms (e.g., K-means) are based on Euclidean distance criteria, which violates exponentially distributed service time assumptions. Thus, the resulting model may not be accurate and may not be able to predict future performance of a modiied system. In order to more accurately extract the assumed exponentially distributed customer class demands from the monitor data, an exponential sieve search technique is proposed and investigated. This technique is based on genetic algorithms (GAs). GAs are search techniques based on the principles of survival of the ttest and genetic recom-bination found in population genetics. GAs have been used successfully in areas as diverse as training an artiicial robot, the traveling salesperson problem, and designing a jet engine to optimize fuel eeciency. Given the success of GAs in these multimodal domains, GAs …