Experimental Evaluation of Software and Systems in Computer Science

This paper proposes a new approach to experimental computer systems research, which we call Literate Experimentation. Conventionally, experimental procedure and writeup are divided into distinct phases: i.e. setup (the method), data collection (the results) and analysis (the evaluation of the results). Our concept of a literate experiment is to have a single, rich, human-generated, text-based description of a particular experiment, from which can be automatically derived: (1) a summary of the experimental setup to include in the paper; (2) a sequence of executable commands to setup a computer platform ready to perform the actual experiment; (3) the experiment itself, executed on this appropriately configured platform; and, (4) a means of generating results tables and graphs from the experimental output, ready for inclusion in the paper.

Our Literate Experimentation style has largely been inspired by Knuth's Literate Programming philosophy. Effectively, a literate experiment is a small step towards the executable paper panacea. In this work, we argue that a literate experimentation approach makes it easier to produce rigorous experimental evaluation papers. We suggest that such papers are more likely to be accepted for publication, due to (a) the imposed uniformity of structure, and (b) the assurance that experimental results are easily reproducible. We present a case study of a prototype literate experiment involving memory management in Jikes RVM.