Workshop Program

Walid Dabbous

Abstract

Scientific evaluation of network protocols requires that experiment results must be reproducible before they can be considered as valid. This is particularly difficult to obtain in the wireless networking domain, where characteristics of wireless channels are known to be variable, unpredictable and hardly controllable. Indeed, anechoic chambers with RF absorbers preventing radio waves reflections and with Faraday cage blocking external interferences represent a ideal environment for experiments reproducibility. However, even if noise and interferences represent unwanted signals, they are unavoidable in real testbeds. It is therefore important to augment the realism of anechoic chamber settings by injecting noise and interfering signals in a controllable way. In this talk we will analyse the difficulties to run reproducible wireless networking experiments in indoor/outdoor environments and describe FIT-R2LAB (Reproducible Research Lab), an anechoic chamber to run realistic and reproducible wireless experiments located at Inria Sophia Antipolis.

Gerhard Wellein

Abstract

Reproducing external performance results or own performance measurements over time is often a big challenge. Even if using the same application code small changes in the deep system software stack (including OS, compiler, BIOS etc) and the dynamic behaviour of modern hardware with its increasing complexity can introduce large measurement changes. This talk will discuss some of these obstacles on modern HPC systems. It will further demonstrate that performance modelling can substantially foster performance reproducibility by providing a clear understanding of the interaction between the application code and the hardware.