Upgrade to Pro — share decks privately, control downloads, hide ads and more …

eBPF Vienna 2025 Presentation

eBPF Vienna 2025 Presentation

Avatar for Teodor J. Podobnik

Teodor J. Podobnik

March 06, 2025

Other Decks in Technology

Transcript

  1. About me • Iʼm Teodor J. Podobnik from Slovenia 󰐶

    • SRE at Prewave • Researcher at Faculty of Computer Science and Information Science, University of Ljubljana • Author of eBPFChirp 2 https://www.prewave.com https://fri.uni-lj.si https://ebpfchirp.substack.com
  2. 4 FaaS (AWS Lambda, Cloud Run, …) Code Execution Container

    Initialization Cold start Warm start Event Deploy Cost vs. Latency
  3. Real-time data analytics Image or video transcoding Web API endpoints

    Demand Constant demand, no cold starts Predictable demand, no cold starts
  4. 6 FaaS (AWS Lambda, Cloud Run, …) Code Execution Container

    Initialization Cold start Warm start Event Deploy
  5. 7 FaaS (AWS Lambda, Cloud Run, …) Code Execution Container

    Initialization Cold start Warm start Event Deploy Network Latency
  6. 9 Client FaaS Provider 50+ ms Source: https://wondernetwork.com/pings 200+ ms

    500+ m s 480+ ms Aprox. FaaS Latency Could we “leverage” network latency to our advantage?
  7. Cloudflare Solution • “Pre-warming” during TLS handshake -> hostname from

    SNI • CloudFlare Workers (NOT containers!) • Not open-source 10 Source: https://blog.cloudflare.com/eliminating-cold-starts-with-cloudflare-workers/
  8. Cloudflare Solution • “Pre-warming” during TLS handshake -> hostname from

    SNI • CloudFlare Workers (NOT containers!) • Not open-source 11 Source: https://blog.cloudflare.com/eliminating-cold-starts-with-cloudflare-workers/
  9. Results 17 • Optimal Speed-Up: • Function “warm start” is

    the limit Network latency >= Function initialization time