Partisan: Enabling Cloud-Scale Erlang Applications.

In this work, we present an alternative distribution layer for Erlang, named Partisan. Partisan is a topology-agnostic distributed programming model and distribution layer that supports several network topologies for different application scenarios: full mesh, peer-to-peer, client-server, and publish-subscribe. Partisan allows application developers to specify the network topology at runtime, rather than encoding topology-specific concerns into application code. Partisan additionally adds support for more channels, enabling users to distribute messages over multiple channels, sometimes in parallel.

We implement and evaluate Partisan in the Erlang programming language and use it in the evaluation of three scenarios. The first scenario compares the raw performance between Distributed Erlang and Partisan, and shows that Partisan performs on par with or better than Distributed Erlang. The second scenario demonstrates that distributing traffic over multiple connections enables Partisan to perform up to 18x better under normal conditions, and up to 30x better in situations with network congestion and high concurrency. The third scenario demonstrates, using existing applications, that configuring the topology at runtime allows applications to perform up to 13.5x better or scale to clusters of thousands of nodes over the general-purpose runtime distribution layer.