Actors and publish/subscribe: an efficient approach to scalable distribution in data centers

Abstract

Data center applications are required to be fault-tolerant and self-healing, and at the same time to scale dynamically with the number of available hardware resources. Highly efficient task distribution is crucial for such services that require low latency and high availability. This paper introduces pub/sub actors as a paradigm to build distributed data center applications without a single point of failure. Our approach does not actively distribute tasks, but uses group communication and an orchestration protocol. Requests are received by a group of potential servers, but only processed by one of them. We present a key-value store using LIBCPPA as a case study of promising performance.