Prof. Dr. Frank Bellosa, Marius Hillenbrand

Abstract:

Main memory bandwidth is a shared resource among applications running on a system. Thus, the behavior of individual applications can have performance implications for other applications running on the same system. This makes main memory bandwidth utilization a interesting characteristic of applications. Furthermore, the patterns of utilization matter: Does an application use a constant amount of bandwidth or does it cause bursts of high utilization separated by phases of low utilization, for example?

As memory accesses are usually invisible to the operating system, measuring the actual memory access behavior of applications in a live system is difficult. In this thesis, we develop a system for capturing main memory bandwidth consumption traces with a high temporal resolution on commodity hardware. To build such a system, we analyze existing hardware mechanisms for directly or indirectly monitoring memory access, such as Intel Memory Bandwidth Monitoring, Performance Counters and Precise Event-Based Sampling (PEBS). We pay special attention to achieving a high resolution while maintaining a low overhead and implement a prototype system based on PEBS.

We evaluate our implementation using both synthetic and realworld benchmarks and find that our system does indeed reduce the incurred worst-case overhead compared to traditional approaches by about 40%. We also find that the reduced overhead of our system increases its accuracy by lowering interference with the measured application.

BibTex:

@mastersthesis{larysch16finegrainedestmembandwitdh,
  author = {Florian Larysch},
  title = {Fine-Grained Estimation of Memory Bandwidth Utilization},
  type = {Master Thesis},
  year = 2016,
  month = mar # "13",
  school = {Operating Systems Group, Karlsruhe Institute of Technology (KIT), Germany}
}