High Speed Active Active clustering over Ethernet RDMA

Existing clustering application works mainly over PCI/FC Protocol, predominantly due to simple to deploy and its suitability for high performance clustering requirements. Recently though, with advent of RDMA technologies (iWARP and RoCE) over High speed Ethernet, Ethernet as a storage interconnect is gaining ground on performance over Fibre channel protocols. This paper talks about various design and implementation considerations to take advantage of low latency and High IOPS RDMA connectivity e.g Lock Less IO command Pool, Pre-allocating CQ/QP memory, Modifications in RDMA Stack, MultiQueue approach to RDMA Ethernet port to make it suitable for clustering application. This paper also talks about challenges for building such applications and other practical considerations, given that clustering applications are different from standard applications and it requires extremely low latency and with minimum spikes. We will also talk about some Active Active implementation topologies and their benefits.