A Replication-based Approach to Parallelizing Protocol Stacks

Current multi-core network protocol stack designs use lock-based and scheduler-based approaches to leverage multiple processors. Unfortunately, these approaches limit protocol stack scalability on multicore systems both for high-bandwidth connections and large numbers of concurrent connections. In this paper, we propose a new approach to parallelizing network protocol stacks based on state replication, as opposed to locking or scheduler-based segregation to dedicated processors. This approach replicates protocol and connection state across processors; each replica is responsible for fully processing only a subset of incoming packets and consistency between these replicas is explicitly managed. This minimizes interprocessor locking and synchronization costs that limit current designs. It also allows for protocol implementations that weaken consistency requirements by using existing protocol-based recovery mechanisms to recover from replica inconsistency. We also describe our approach to implementing this protocol stack architecture and discuss the open research questions this approach raises to network stack and protocol design.