The maintainers of Linkerd have announced the results of the latest benchmark analysis of service mesh performance. These tests, designed to measure mesh performance under real-world conditions, showed that Linkerd consumed an order of magnitude less memory and CPU and introduced up to 400% less latency than Istio at scale.

The benchmarks evaluated the latest versions of Linkerd and Istio and reported the relative costs of each mesh against the baseline condition of no service mesh. These costs included both the CPU and memory consumed by the mesh as well as the user-facing latency the meshes introduced.

At the highest traffic level measured, the benchmarks showed:

• Latency: Istio introduced 36% more additional median latency and 438% more additional maximum latency than Linkerd.
• Control plane: Istio’s control plane required 2.5x the memory and over 50x the CPU that Linkerd’s control plane required.
• Data plane: Istio’s Envoy proxies consumed 8x the memory and 8x the CPU that Linkerd’s Rust-based proxies did.

The benchmark tests followed a 2019 analysis performed by Kinvolk which showed a similar performance gap between the meshes.

Today’s tests use the same Kinvolk benchmark suite with latest stable releases of both projects (Linkerd 2.10.2 and Istio 1.10.0) and run on a CNCF Kubernetes cluster provided by Equinix Metal. While the 2019 Kinvolk tests evaluated performance at 500 requests per second (RPS) and 600 RPS, these tests evaluate the service meshes at over three times the scale at 2,000 RPS.