Hello,
since issue #8938 was closed, posting here. I’ve ran some experiments (this time using Istio incl. gateway, not just Envoy) and summarized the results in this report.
TLDR is that I got decent results even with 2 CPUs for sidecar (6 or 14 other CPUs were dedicated to application), but using more worker threads (--concurrency) than #CPUs gave me better results.