Weekly Notes 42: Watched a couple more p99conf talks and read about how Cloudflare distributes traffic across datacenters in its network

• Architectural, operational, organizational challenges often found in systems and teams

• Dreamworks render farm presentation at Ubuntu Summit 2024: Running a large cluster of computers to render frames in a dreamworks movie

• How Cloudflare’s systems dynamically route traffic across the globe: Make network routing changes for traffic between datacenters automatically to maintain a target latency for users. The system has to be aware of what capacity is available in a datacenter and neighbouring ones to make these decisions.

• My approach to GTD and PKM: I like the 3 risks he calls out that he’s worried about. Seems relevant to many of us:

  • Lots of things going on so I forget things
  • Lots of things in my head that I can’t easily express in code or docs or elsewhere but that would be helpful if shared
  • Trap: Focus on short term easy wins over longer term harder work

p99 conf 2024

• Zero-overhead Container Networking with eBPF and Netkit: Amazing work to bring container network performance (for high line rates eg 100Gb) in line with host networking. There used to be a virtual network adapter inside a container paired with one outside and 2 full network stacks - one inside and one outside the container. In kernel 6.6 egress traffic bypasses the host network stack with the help of ebpf programs and new netkit adapters. Neat stuff! :)

• Enhancing P99 Latency: Strategies for Doubling/Tripling Performance in Third-Party APIs: TLS1.3 for fewer roundtrips during connection establishment, http2 for connection multiplexing and async dns lookups. Told a few stories about the kinds of problems they encountered communicating with 3rd party apis @ uber. With enough api usage, small-ish sounding things can make a big difference. (Issue amplification by volume)

• Garbage Collection in Java: The progress since JDK 8 by Stefan Johansson: Compares gc latency, memory usage, gains in efficiency going from java8 through to java21. Great talk. Lots of improvement in the last 10 years or so.