Granular CPU Capacity Management at Scale with eBPF

Real-time market data is exceptionally bursty, with update rates in the busiest seconds of the day regularly exceeding 10x the average. User experience is predicated on maintaining sufficient CPU headroom to prevent full buffers and the resulting client disconnects. Sampling cumulative CPU time at a typical scrape interval hides microbursts, and sub-second polling from user space induces unacceptable overhead, so a different approach is needed.

This talk will cover how Market Data SRE at Goldman Sachs uplifted CPU monitoring of our market data distribution infrastructure in an unintrusive way, achieving 10x the granularity with 5% of the original monitoring overhead. We will cover the journey from deciding to use eBPF, through trials using bpftrace and making the leap to BPF C, to collecting and aggregating the metrics effectively. It will be most relevant to those interested in capacity management across a heterogeneous estate, and those looking to implement eBPF for the first time in their organisations.