Fast NIC based RDMA implementation for adaptive unreliable networks

Remote Direct Memory Access (RDMA) is one of the basic communication fabrics for parallel computers. Its performance is crucial for most parallel workloads. For next generation super computers, the system usually extends to extremely large scale with millions of computation cores connected through the interconnection network. Most interconnection network features out of order packets delivery and low system-wide network reliability. Building high performance and reliable point-to-point communication over next generation interconnection network is a challenging task. Current system usually implements RDMA through two approaches: 1) the receiver side counters approach; 2) sender side sliding window approach. We see that current approach works better on reliable interconnection networks, but has obvious performance degradation over the un-reliable network. This paper proposes a new hardware approach for fast RDMA transmission, which can provide scalable performance on un-reliable network. Our approach uses a novel receiver side sliding window to support out-of-order packets delivery. Our approach proposes a new implementation of the traditional sliding window approach, which uses receiver side sliding window to efficiently retransmit partial RDMA data when part of the packets fail to reach the receiver. The experiments show that for low reliability interconnection network, our approach has scalable performance benefit over current RDMA implementations.