Unified high-performance I/O: One stack to rule them all

Abstract

Fast non-volatile memories are exposing inefficiencies in traditional I/O stacks. Though there have been fragmented efforts to deal with the issues, there is a pressing need for a high-performance storage stack. Interestingly, 20 years ago, networks were faced with similar challenges, which led to the development of concepts and implementations of multiple high-performance network stacks. In this paper we draw parallels to illustrate synergies between high-performance storage requirements and concepts from the networking space. We identify common high-performance I/O properties and recent efforts in storage to achieve those properties. Instead of reinventing the performance wheel, we advocate a case for using mature high-performance networking abstractions and frameworks to meet the storage demands, and discuss opportunities and challenges that arise with this unification.