Skip to content

Istio Performance

carolynhu edited this page Aug 26, 2020 · 8 revisions

Istio performance is measured in several different ways.

  1. IBM Regpatrol measures performance of a realistic application. It runs nightly in ingress-only, full-istio and mixer-disabled modes. The throughput and response times can be compared between builds to detect trends and regressions.

  2. Synthetic service graph measures performance of a deeply nested synthetic micro-services application. It creates 20 services per name space and typically runs with 20 - 40 namespaces in large a cluster. Here is a v1.1 Performance dashboard and the corresponding Mesh dashboard using release-1.1-20190110-09-15 build.

  3. Two pods benchmark test runs a focused benchmark for data path performance only, which is available at: https://perf.dashboard.istio.io/

performance numbers

release-1.1-20190110-09-15

sidecars - 0.75 vCPU / 1k rps @1KB response size

istio-telemetry - 0.7 vCPU / 1k rps

istio-policy - 0.002 vCPU / 1k rps @ > 95% cache hit rate

twopod test baseline: {p90 1.2ms, p99 2.1ms}, full istio: {p90 9.6ms, p99 13.8ms} Istio adds 8ms at p90 and 11ms at p99.

Pilot performance 1.1-RC2

Var experiment 1 experiment 2 experiment 3 experiment 4
Pilots 1 1 2 2
Sidecars 1600 2000 3000 1600
Services 800 1000 1500 800
VirtualServices 80 100 150 80
Gateways 44 54 81 44
Namespaces 40 50 75 40
Config Change Rate 10/m 12/m 18/m 10 /m
Pilot VCPU peak 1 1 1.5 5
Pilot Mem 1.4GB 1.4GB 1.8GB 4GB
Sidecar Mem 70MB 80MB 85MB 170MB
Namespace Isolation True True True False
  • Pilot: 1
  • Sidecars: 1600
  • Services: 800
  • VirtualServices: 80
  • Gateways: 44
  • Namespaces: 40
  • Config Change rate: 10 per minute {galley.sourceEventRate}

Namespace Isolation: On, Sidecar resource

  • 1 VCPU peak

  • 1.4GB Memory

  • Proxy Memory: 70Mi

Namespace Isolation: off

  • 5 VCPU peak
  • 4GB Memory
  • Proxy memory: 170Mi

When NS isolation is off

  1. Pilot needs to broadcast changes to all pods and namespaces
  2. Every proxy gets more configuration. All clusters and all listeners are delivered to all sidecars. That explains the difference between memory consumption.

Time to get ready. < 5s

Dev Environment

Writing Code

Pull Requests

Testing

Performance

Releases

Misc

Central Istiod

Security

Mixer

Pilot

Telemetry

Clone this wiki locally