Decreasing memory footprints for better enterprise Java application performance

Abstract

In this paper, we present a work for reducing the memory footprint of enterprise Java applications. The work relies on the predictions provided by stochastic models of the applications' data-access patterns. The models, built during the execution of the application, are used both at compile-time, to control the in-memory representation of data, and, at runtime, to decide which portions of the data to load. The combined effect of these two approaches allows for an effective reduction in the memory used by the application, leading to a significant performance improvement. We evaluate the newly developed approaches on the TPC-W benchmark, with different database sizes, and show that our solution increases the benchmark throughput by 10.78% on average, with a maximum of 35.43% when operating over larger databases. © 2012 Springer-Verlag.