Core-Local Multiplexing for Accelerators in M³

Recent system designs increasingly rely on various specialized accelerators that optimize specific workloads with high performance and low power consumption. M³ is an operating system that improves integration of accelerators by adding a uniform communication interface that can be used by accelerators to access OS services like the file system.

M³ also allows applications to share accelerators with minimal hardware support by triggering context switches remotely from the shared M³ kernel. Unfortunately, this means that the shared kernel quickly becomes the limiting factor for context switch performance in the entire system.

In my thesis I investigated how context switches can be performed locally for a simple SHA-3 hash accelerator (without involving the M³ kernel). The idea is to add a simple CPU core next to the accelerator that provides flexibility to implement scheduling and service protocols, while maintaining the performance and efficiency of the accelerator.

(Beleg)