A memory access detection methodology for accurate workload characterization

Tools for memory access detection are widely used, playing an important role especially in real-time systems. For example, on multi-core platforms, the problem of co-scheduling CPU and memory resources with hard real-time constraints requires a deep understanding of the memory access patterns of the deployed taskset. While code execution flow can be analyzed by considering the control-flow graph and reasoning in terms of basic blocks, a similar approach cannot apply to data accesses. In this paper, we propose MadT, a tool that uses a novel mechanism to perform memory access detection of general purpose applications. MadT does not perform binary instrumentation and always executes application code natively on the platform. Hence it can operate entirely in user-space without sand-boxing the task under analysis. Furthermore, MadT provides detailed symbolic information about the accessed memory structures, so it is able to translate the virtual addresses to their original symbolic variable names. Finally, it requires no modifications to application source code. The proposed methodology relies on existing OS-level capabilities. In this paper, we describe how MadT has been implemented on commercial hardware and compare its performance with state-of-the-art software techniques for memory access detection.