Towards Asynchronous System Calls for Concurrency Platforms
Florian Schmaus and Florian Fischer	

A Concurrency Platform (CP) consists of a parallel programming language and a runtime-system transforming the programming language's concurrency constructs into parallelism. The control-flow abstractions from this transformation emerging are called fibers. Typically, a CP's runtime system will use kernel-level threads to execute those fibers, employing an m:n threading model. Unfortunately, if a fiber executes a blocking system call, its carrying worker-thread will also block, usually causing underutilization due to increased off-CPU time. Fortunately, modern operating systems are in the progress of establishing asynchronous system call interfaces. We show how concurrency platforms can efficiently use these interfaces to increase application throughput and to reduce latency. Furthermore, we identify the new challenges asynchronous system-call interfaces pose to an operating system's architecture.