4 more exciting talks accepted by the illustrious PC.
1) Ido Schimmel and Petr Machata will discuss the kernel 5.13 feature supporting resilient nexthop groups. Resilient NH groups are intended to minimize disruption in flow routing when changes are made to the group composition and weights of constituent nexthops. This talk will detail the inner workings of resilient nexthop groups, including hardware offload. The talk will also cover future work such as static resilient nexthop groups, Nexthop group consolidation, and integration into FRR.
2)Kostis Kaffes, Jack Tigar Humphries, David Mazières, and Christos Kozyrakis address the notion that OS schedulers often, if not always, have discrepancy in their algorithms that do not take care of what networked applications need because they lack system level insight into those needs. There is a lack of coherency in the kernel between the scheduling and the various layers, including the network stack, NICs, and applications. The result is suboptimal scheduling decisions which are often solely responsible for applications poor latency and throughput. Kostis et al propose "Maple" - a framework for user-defined scheduling. A user expresses an application-specific scheduling policy in a few lines of ebpf code and then deploys it across system layers without modifying their code. Kostis et al claim up 8× performance improvement compared with default policies.
3) Mark Claypool, Benjamin Peters, Pinhan Zhao, and Jae Won Chung evaluate TCP Hystart performance on satellite links. To avoid overshooting TCP slow start ramp up, the Linux kernel has a Hystart mode (on by default) that will exit slow start before packets are lost. Mark et al demonstrate that Hystart-Delay often causes spurious exit of slow-start on the satellite links they tested. This has big implications for efficient utilization of commercial satellite Internet links. Mark et al will be seeking suggestions from the community on how to overcome these challenges. For the moment, they have a suggested configuration which is being used as a short term solution. They will also briefly discuss future work to address the issue.
4) Eric Dumazet is not a man who uses many words. The title of his talk is simply "Big TCP". In this talk he discusses challenges faced when the network stack now has to deal with 200/400Gbit speeds. How to address this? Once one reduces or avoids use<->kernel copying the size of TSO/GRO packets matters. In this talk Eric will present how we can override current limits and reduce TCP/IP stacks overhead.
cheers, jamal