Are You Sure About Your Mesh Performance?
summary
Benchmarks on Envoy use bare metal tuned boxes, VMs with DPDK, and sandboxes; Fortio, wrk and Nighthawk report wildly different latencies because of connection reuse, request-timing jitter and closed vs open loop behavior. Four cores handle ~1,000 TPS but 48 cores push a 10,000 TPS input to >7,000 resolved TPS and reduce p99 by ~4.5x; adding Istio in VMs multiplies latency by ~3–4x and cuts throughput by ~30–60% at 10k TPS. Goal: >50k QPS and <5 ms with CNI+proxy and hardware offloads.
AI-written summary · about
chapters (5)
- 0:00Introduction and methodology
- 2:29Bare metal performance study
- 5:50Virtualized environment tests
- 11:32Benchmarking pitfalls
- 23:22Future work and optimization
related talks
40mAchieving the Ultimate Performance with KVM
27mA Service Mesh Benchmark You Can Trust
09mBenchmarking 5 Popular Ingress Controllers
43mRealtime performance at scale (...)
18mPower Savings & NFV: How to reconcile these 2 constraints ?
36mFine Grained Mesh Metrics for Better Visibility With Native Performance
26mRunning Envoy at the Edge
34mIstio's Ambient Mesh: The Real Cost of Sidecar-less Tracing