A neural network approach to forecasting computing-resource exhaustion with workload

Software aging refers to the phenomenon that applications will show growing failure rate or performance degradation after longtime execution. It is reported that this phenomenon usually has close relationship with computing-resource exhaustion. This paper analyzes computing-resource usage data collected on a LAN, and quantitatively investigates the relationship between computing-resource exhaustion trend and workload. First, we discuss the definition of workload, and then a Multi-Layer Back propagation neural network is trained to construct the nonlinear relationship between input (workload) and output (computing-resource usage). Then we use the trained neural network to forecast the computing-resource usage, i.e., free memory and used swap, with workload as its input. Finally, the results were benchmarked against those obtained without regard to influence of workload reported in the literatures, such as non-parametric statistical techniques or parametric time series models.