Sol Boucher, Anuj Kalia, and David G. Andersen, Michael Kaminsky*
Carnegie Mellon University
Modern cloud computing environments strive to provide users with fine-grained scheduling and accounting, as well as seamless scalability. The most recent face to this trend is the “serverless” model, in which individual functions, or microservices, are executed on demand. Popular implementations of this model, however, operate at a relatively coarse granularity, occupying resources for minutes at a time and requiring hundreds of milliseconds for a cold launch. In this paper, we describe a novel design for providing “functions as a service” (FaaS) that attempts to be truly micro: cold launch times in microseconds that enable even finer-grained resource accounting and support latency-critical applications. Our proposal is to eschew much of the traditional serverless infrastructure in favor of language-based isolation. The result is microsecond-granularity launch latency, and microsecondscale preemptive scheduling using high-precision timers.
FULL PAPER: pdf