Virtual reuse distance analysis of SPECjvm2008 data locality

Reuse distance analysis has been proved promising in evaluating and predicting data locality for programs written in Fortran or C/C++. But its effect has not been examined for applications in managed runtime environments, where there is no concept of memory address. For this reason, traditional reuse distance analysis based on memory addresses is not directly applicable to these applications. This paper presents the Virtual Reuse Distance Analysis (ViRDA), which resolves the difficulties associated with runtime environments and provides insights into the highlevel locality in dynamic applications. ViRDA addresses the problem caused by managed runtime artifacts, garbage collection in particular, by using virtual data identities, obtained through a standard profiling interface, to capture inherent data locality. The effectiveness of ViRDA is evaluated using a subset of the SPECjvm2008 benchmark suite. The new analysis reveals the reuse distance signatures of these programs and helps to explain the cause of excessive cache misses. It also predicts locality for large inputs based on training analysis of several small inputs. The prediction error is no more than 6% for the 4 scimark workloads.