The End of slow networks: It's time for a redesign

The next generation of high-performance networks with re-mote direct memory access (RDMA) capabilities requiresa fundamental rethinking of the design of distributed in-memory DBMSs. These systems are commonly built underthe assumption that the network is the primary bottleneckand should be avoided at all costs, but this assumption nolonger holds. For instance, with InniBand FDR 4×, thebandwidth available to transfer data across the network isin the same ballpark as the bandwidth of one memory chan-nel. Moreover, RDMA transfer latencies continue to rapidlyimprove as well. In this paper, werst argue that traditionaldistributed DBMS architectures cannot take full advantageof high-performance networks and suggest a new architec-ture to address this problem. Then, we discuss initial resultsfrom a prototype implementation of our proposed architec-ture for OLTP and OLAP, showing remarkable performanceimprovements over existing designs.