Pitfalls of Asynchronous Operating Systems for Many-Cores

Authors: Randolf Rotta, Robert Kuban, Jörg Nolte

The arrival of many-core architectures with hundreds of slow hardware threads motivates the redesign of operating system kernels and the overall system management stack in order to actually achieve high throughput and fast execution.  It seems reasonable to incorporate parallel processing of compute-heavy management tasks, to improve data locality by centralising services and sending tasks to them, to deffer less important tasks to idle time or idle threads, and to fill waiting times with other pending tasks instead of spinning on locks. 
Asynchronous execution models and message-based coordination are a natural choice to meet these goals as is demonstrated by distributed designs such as Barrelfish, fos and many more.  However, some of the goals require dynamic load balancing and, thus, sharing. Bringing together message passing and shared memory combines the disadvantages of both worlds -- leading to a number of pitfalls for the OS design.
The presentation discusses some of these pitfalls and outlines possible solutions based on our experiences from the ongoing MyThOS project. Many problems are not new but need to be revisited in their specific combination and in the many-core context. We hope to stimulate a wider discussion about the implications of asynchronous operating systems.