Section: New Results

Resource Allocation in Large Data Centres

Participants : Christine Fricker, Philippe Robert, Guilherme Thompson, Veronica Quintuna Rodriguez.

With the emergence of new networking paradigms such as Cloud Computing and related technologies (Fog Computing, VNF, etc.) new challenges in understanding, modelling and improving systems relying on these technologies arise. Our research goal is to understand how the stochastic nature of the access to these systems affects their performance, and to design algorithms which can improve global performance using local information. This research is made in collaboration with Fabrice Guillemin, from Orange Labs.

Building up from the results previously obtained by this team, we have extend our research towards more complex systems, investigating the behaviour of multi-resource systems, which are globally stable but local congested, a problem that naturally arises from the decentralization of resources. We investigate a cooperation scheme between processing facilities, where congestion-maker clients, the one with the largest demand the locally congested resource are systematically forwarded to the another data centre when some threshold on the occupation level is reached. These thresholds are chosen to anticipate sufficiently in advance potential shortages of any resource in any data centre. After providing some convergence results, we are able to express the performance of the system in terms of the invariant distribution of an inhomogeneous random walk on the plane. We derive optimal threshold parameters, improving the performance of the distributed Cloud Computing system in such a way that it approaches the efficiency of a centralised system. Currently, a document is being prepared for publication, but the main results are presented in G. Thompson's PhD Document [2].