By Alexander Konovalov, Research Fellow, Centre for Interdisciplinary Research in Computational Algebra, University of St Andrews
1. The talk about the Recomputation.org project was very enthusiastically received by the audience.
2. Some interesting questions and suggestions we made (see event report for their outline).
3. Some members of Scottish Programming Languages Seminar community may be interested in contributing recomputable experiments to Recomputation.org when we will be able to offer the necessary infrastructure.
The title of my talk was "Recomputation in Scientific Experiments" (slides and tweet). I've collected the following notes from the communication at SPLS:
1) There is an agreement that the following issues are very important:
- - long term virtual machine persistence;
- - availability of automatic configuration tools to reduce the bandwidth;
- - security: what if VM was used by someone for a long time and then uploaded? It may have confidential (personal or corporate) data. Also, how we are going to trust to users uploading theirs VM?
2) Is it possible to connect several VMs together? For example:
- - to be able to quickly fire up virtual clusters;
- - to do pipelining when data produced on one VM need to be transferred to another VM (e.g. two or more software components needed for the experiment only work in some particular Linux distributions and can not be combined in one VM).
3) Part of the SPLS community are users of various theorem provers. This this one of the target audiences for Recomputation.org. They are less interested in reproducing performance, but mostly interested in artefacts produced by computational experiments.
4) Regarding the performance, indeed we state that the runtime is secondary. What could be done probably is the relative runtime - e.g. comparing old and new version of an algorithm and observing that the new one is ~N times faster. That might also depend on the architecture where the VM is running, though, and may be an interesting research problem.