Scalable packet processing with many queues and cores

Session information has not yet been published for this event.


One Line Summary

Advancements in NIC hardware and functionality, as well as things like increasing core counts on systems present new opportunities for taking networking performance to next level.


Network interfaces with multi-queue, flow-aware load balancing (aRFS) and protocol filtering (ntuple) can intelligently distribute load on multiprocessors. The fanout allows host packet processing to shift from a multi-stage interrupt-driven ‘vertical’ stack to concurrent ‘horizontal’ pools of polling threads. Initial benchmarks show a reduction in kernel latency by 5-fold to below 10 μs. This talk reviews the new hardware features, shows preliminary results and presents a prototype Linux implementation of horizontal processing.


performance, networking, NIC, scalability


  • Biography

    Willem is a kernel Engineer at Google, working mainly on scaling packet processing in the network stack.

    Before joining Google, he co-developed the Nexus trustworthy microkernel at Cornell and designed various network processing systems at the Vrije Universiteit Amsterdam, such as pipesfs: a virtual filesystem for interacting with live kernel I/O using memory mapped descriptor rings and packet buffers.