Apt: A Platform for Repeatable Research in Computer Science

Repeating research in computer science requires more than just code and data: it requires an appropriate environment in which to run experiments. In some cases, this environ-ment appears fairly straightforward: it consists of a par-ticular operating system and set of required libraries. In many cases, however, it is considerably more complex: the execution environment may be an entire network, may in-volve complex and fragile configuration of the dependencies, or may require large amounts of resources in terms of com-putation cycles, network bandwidth, or storage. Even the “straightforward ” case turns out to be surprisingly intricate: there may be explicit or hidden dependencies on compilers, kernel quirks, details of the ISA, etc. The result is that when one tries to repeat published results, creating an envi-ronment sufficiently similar to one in which the experiment was originally run can be troublesome; this problem only gets worse as time passes. What the computer science com-munity needs, then, are environments that have the explicit goal of enabling repeatable research. This paper outlines the problem of repeatable research environments, presents a set of requirements for such environments, and describes one facility that attempts to address them. 1