Scalability and Sizing Guide
Setup multiple replicas of the control plane components.
Split mixer check and report pods.
High availability (HA).
See also Istio’s Performance oriented FAQ
And the Performance and Scalability Working Group work.
Current recommendations (when using all Istio features):
1 vCPU per peak thousand requests per second for the sidecar(s) with access logging (which is on by default) and 0.5 without,
fluentdon the node is a big contributor to that cost as it captures and uploads logs.
Assuming typical cache hit ratio (>80%) for mixer checks: 0.5 vCPU per peak thousand requests per second for the mixer pods.
Latency cost/overhead is approximately 10 millisecond for service-to-service (2 proxies involved, mixer telemetry and checks) as of 0.7.1, we expect to bring this down to a low single digit ms.
mTLS costs are negligible on AES-NI capable hardware in terms of both CPU and latency.
We plan on providing more granular guidance for customers adopting Istio “A la carte”.
The goal for 2018 for Istio is to reduce both the CPU overhead and latency of adding Istio to your application but please note that if you application is handling its own telemetry, policy, security, network routing, a/b testing, etc… all that code and cost can be removed and that should offset most if not all of the Istio overhead.