Method and system for parallelizing completion event processing

Abstract

Disclosed are methods for handling RDMA connections carried over packet stream connections. In one aspect, I/O completion events are distributed among a number of processors in a multi-processor computing device, eliminating processing bottlenecks. For each processor that will accept I/O completion events, at least one completion queue is created. When an I/O completion event is received on one of the completion queues, the processor associated with that queue processes the event. In a second aspect, semantics of the interactions among a packet stream handler, an RDMA layer, and an RNIC are defined to control RDMA closures and thus to avoid implementation errors. In a third aspect, semantics are defined for transferring an existing packet stream connection into RDMA mode while avoiding possible race conditions. The resulting RNIC architecture is simpler than is traditional because the RNIC never needs to process both streaming messages and RDMA-mode traffic at the same time.