Please log in to watch this conference skillscast.
Running compute clusters at a low hardware utilization is wasteful. Oversubscription of resources promises to increase utilization, but can be pretty challenging in the presence of tight latency objectives. Fortunately, there seems to be a feasible middle ground.
In this talk, Stephan will explore the Apache Mesos oversubscription primitives and how they leverage them in a simple threshold-based solution. Stephan will explain why this approach works for him and his team and what they have learned so far. In order to provide sufficient context, the talk will also shortly discuss Linux cgroups and relevant service-level monitoring signals.
As their code is open-source, Stephan and his team hope that others can immediately profit from their work and cut down your hardware spending as well.