Bridging the Performance Gap Between Converged RDMA Dataplane and Kernel-Bypass

While kernel-bypass is often considered an essential feature of high performance networks, it breaks the traditional OS architecture and prevents a fine-grain control of the kernel over the communications. Converged RDMA Dataplane (CoRD) is an attempt at solving this problem by giving back the control over dataplane operations to the kernel effectively enabling a better control and utilisation of the network resource. However, CoRD suffers from a massive overhead compared to its kernel-bypass counter-part. CoRD was implemented by porting the user library code inside the Infiniband driver and thus their performance should be equivalent but our measurements show that the overhead exceeds the delay imposed by a system call. In this work, we identified the causes of this overhead and improved the driver performance by up to 65% to reach a sub-microsecond overhead compared to kernel-bypass.

(Master defense)

(Online meeting only)