OCPSummit19 - EW: Networking: Hardware - DCTCP in the OCP Datacenter Utilizing OCP Hardware

Over the past year much work has been done to tie in the DCTCP congestion control algorithm with OCP networking [1] and server [2] hardware. This has helped solve numerous problems related to tail latencies. This talk will describe a method to utilize this work in adopting disaggregated flash storage over TCP within a low power OCP environment focused on the Wedge100 switch and the Yosemite v2 multi-host platform. It will describe the OCP specific operational challenges faced, solutions found, and the impact this has with respect to disaggregated data center operation. Finally, the interaction between the bottlenecks addressed in the Wedge100 rack switches and in the Yosemite platforms will be explored, including how through the combination of queueing structures in the switches and OCP v2.0 mezzanine cards, tail latencies can be reliably capped. Scope DC network switches, disaggregated flash workloads, storage microbenchmarks, latency sensitive DC apps, and NIC dataplane behavior. Presented by Lawrence Brakmo ,Nicolaas Viljoen ,Amal Tariq