Open vSwitch 2016 Fall Conference

When service providers evaluate virtualization of their network infrastructure, one of the key areas of focus is TCP performance, a critical primitive for many of their end-to-end network services.

This talk describes an approach for optimizing throughput of TCP-based workloads in an Open vSwitch with DPDK (OvS-DPDK) setup. Our method involves implementing support for DPDK’s TCP Segmentation Offload (TSO) feature within OvS-DPDK as the primary performance booster. We additionally describe a tuned software configuration that, in concert with DPDK’s TSO and vHost multi-queue features, has demonstrated optimal performance for TCP workloads. Finally, we outline how to implement an OvS-based virtualized TCP workload test server, and three key test scenarios that reflect real-world workloads.

In an analysis of the test results, we found that our accelerated virtualized broadband speed test server could process external requests with an average throughput of 9.35Gbps over a 10Gbps network connection. In addition, the same virtualized server was able to process internal workloads with an average throughput of 40Gbps. These figures constitute performance gains of 2.4x and 10x, respectively over baseline figures. Consequently, we conclude that the TCP performance optimizations we implemented result in an optimal NFVI capable of handling communications-grade Network Functions Virtualization (NFV) TCP workloads in a real-world deployment. This was all made possible with the close collaboration between Intel and one of its flagship customers.