SCHEDFAST: Understanding Linux scheduling bottlenecks in large scale video streaming contexts

Context. 

With the emergence of the covid-19 pandemic, we are witnessing as a consequence a rise of large-scale video conferencing that appears as an adequate solution to connect people together in spite of the quarantine. Specifically, video conferencing platforms have been expected to become an increasingly popular alternative to in-person events, and hence become a corner stones for maintaining any kinds of remote activities. The robustness, in terms of scalability and efficiency, of such systems has been put under high stress conditions with their increasing popularity. One solution commonly adopted to deal with this kind of emergency situations is to oversize the resources allocated to these systems. For instance, many providers such as Microsoft, Zoom, Discord, Slack, as well as our academic institutions such as INRIA and/or CNRS, have blindly allocated a huge amount of hardware resources to cope with their underlying saturation. If this naive approach is understandable and successful in an emergency situation, it however cannot be envisaged in the long term, given that the end of the pandemic is not yet in a close sight, and the inherent cost of such solutions. As far as our knowledge there is not yet research on understanding bottlenecks in such systems as it is the first time that it has such a huge success

Objectives. 

The main aim of the PhD is to understand the behavior of the multicore Linux Scheduler under a heavy load [1,2,3,4] such as the one observed during videoconferencing sessions. Particularly, our aim is to study how to characterize the type of load that the scheduler has to managed, through the use of statistical models such as for instance Markov chain models, to then to study how we can define/optimize heuristics of the scheduler according to the load it faces to make it scales better.