BSC is pioneering use of Extreme Value Theory in Computer Sciences

Monday, December 14, 2015

Thread Assignment in Multicore/Multithreaded Processors: A Statistical Approach, was selected to be the featured article of the January 2016 issue of IEEE Transaction on Computers (TC). From January, the article will appear in Computing Now , and will be promoted by TC editorial through videos in English and Chinese.

In the study, experts from BSC Computer Sciences department use a statistical approach to the problem of thread assignment, which is a part of process scheduling on modern processors. The authors use random sampling and Extreme Value Theory, branch of statistics that estimates extreme population values, to predict the performance of the optimal thread assignment. They also show that, if no suitable heuristic based algorithm is available, a sample of several thousand random thread assignments is enough to obtain, with high confidence, an assignment with performance close to optimal.

In addition to the previous work on the thread assignment problem, BSC researchers have applied random sampling and EVT for graph partitioning of streaming applications. Also, BSC experts are leading the European project called PROXIMA, that pretends to apply these techniques to the analysis of applications running in real-time systems.

EVT is an important branch of statistics, which has found multiple applications in civil engineering, material testing, finance and risk management. Its application in computer science is, however, currently marginal. BSC researchers from Computer sciences department believe that EVT provides many powerful theorems and tools, which could be of great interest to our community.

Therefore, BSC is pioneering use of EVT in computer science and engineering being sure that our successful experiences will encourage new and fruitful applications of EVT in these fields.