Imagine you have a supercomputer in your house but don't know about it. What sounds like an abstract thought experiment is reality for many companies which leave the enormous amount of idle CPU cycles of their employees' computers unused. There are many understandable reasons why not to run mission critical business processes on end-user controlled commodity hardware, including security, privacy, deadlines, and the omnipresent risk of failures. We argue that with the right programming abstraction and a context-aware task scheduler, a heterogeneous and unreliable pool of end-user devices can turn into a responsive and reliable computing cluster.

In this talk, I will present the Tasklet system, our approach for a decentralized task offloading architecture. I will provide an overview of several optimization strategies and heuristics that we in the group of Distributed Operating Systems (DOS) use to overcome the shortcomings of end-user devices as offloading targets. Core to these approaches is a context-aware task-scheduler that decides whether, how, and whereto offload computational work.