Networks are getting faster but so is storage. It is not uncommon anymore for a single high performance storage device to deliver as much as a million read/write operations per second. For traditional remote storage accesses CPU requirements have become unsustainable. The recently merged NVMe-over-TCP has improved performance lowering CPU utilization. However, despite of its advantages of being able to use commodity NICs, NVMe-over-TCP's performance is still significantly worse than NVMe-over-RDMA.
Rachit Agarwal, Qizhe Cai, and Jaehyun Hwang have some new ideas on how to significantly improve the status quo. They will discuss a design, implementation and evaluation of several new ideas within NVMe-over-TCP. Their design maintains all the desirable properties of NVMe-over-TCP (unmodified applications, unmodified TCP/IP stack, etc.), and yet, saturates a 100Gbps link for remote accesses using CPU utilization _similar to state-of-the-art user-space and RDMA-based solutions_.
More information: https://bit.ly/2Ri6hJd
cheers, jamal