Improving Performance in HPC Centers by Modeling Users Through an Evolutionary Virtual Interface

An agent-based profiling approach is presented in this paper in the form of a virtual interface that models user behavior and satisfaction with the objective of improving the performance of high-performance computing (HPC) centers. The interface´s function is to translate user requests and satisfaction criteria into what is really necessary, thus permitting the optimization of the HPC center´s scheduling, taking into account the predicted user satisfaction together with the objectives of the management of the center in terms of resource usage and cost. The system is built using an evolutionary agent-based profiling architecture where agents are evolved in real time to adapt to the different users. These agents cooperate with the scheduling mechanism, providing resource usage estimations for the different tasks, as well as predicting the effect of possible strategies on user satisfaction. This paper focuses on the user behavior modeling component, although the global architecture is also presented. Some experiments are carried out where the proposed architecture interacts with a real job-management system (JMS) (sun grid engine). They clearly show that modeling the user and taking user satisfaction into account helps to improve system performance.