TLS1.3 performance problem

[zh-jq-b]

See https://bytedance.larkoffice.com/docx/NEqXdDyq8op8mwx0XTWcwgionnf.

The tls1.3 performance isn't as good as OpenSSL 1.1.x when testing at high concurrency.

[djc]

Why are you not testing aws-lc-rs support? Are your results in line with https://github.com/aochagavia/rustls-bench-results?

[zh-jq-b]

Why are you not testing aws-lc-rs support?

The test is against rustls 0.21 branch and I will try 0.22 ring & 0.22 aws-lc in the next few days.

We did test 0.22 here https://bytedance.larkoffice.com/docx/TSqmdLNIyozJdCx4rzqcjPyonjd, and the difference between 0.21&0.22ring&0.22aws-lc is not too much.

Are your results in line with https://github.com/aochagavia/rustls-bench-results?

No, I haven't checked that.

[zh-jq-b]

I also add tls1.2 test result here https://bytedance.larkoffice.com/docx/NEqXdDyq8op8mwx0XTWcwgionnf.
The rustls 1.2 performance is better than tls1.3 & openssl (which is expected).
There are no much RPS difference between rustls 0.21 & rustls 0.22 ring & rustls 0.22 aws-lc in our test.

[djc]

There have been some recent changes on the main branch. Would it be hard to try this with current rustls main?

[zh-jq-b]

There have been some recent changes on the main branch. Would it be hard to try this with current rustls main?

The max RPS (ring) is still around 54000.

I just bump the version here bytedance/g3@0c2feaf without any API changes.

[cpu]

Are you building with gcc or clang?

[zh-jq-b]

Are you building with gcc or clang?

I didn't set a custom CC, so it should be gcc.

[briansmith]

I didn't set a custom CC, so it should be gcc.

recent clang should be faster.