pq-code-package / mlkem-native

High-assurance, high-performance ML-KEM implementation for mobile, pc, and server targets
https://pq-code-package.github.io/mlkem-native/dev/bench/
Apache License 2.0
11 stars 9 forks source link

Benchmark c7i on metal #447

Closed hanno-becker closed 6 hours ago

hanno-becker commented 6 hours ago

Current benchmarks on c7i.large exhibit a high performance variability that is likely due to virtualization.

This commit switches to a metal instance for c7i to counter this problem.

https://pq-code-package.github.io/mlkem-native/dev/bench/ suggests that the issue does not appear for c6a, c6i, c7a, so those are not touched (and indeed it would be non-trivial to do so because we'd exceed the vCPU quote with more than one metal instance).