AW: AW: RE: slow https performance compared to http

Lukas Tribus luky-37 at
Mon Nov 14 08:57:39 UTC 2016

> I agree but I think that separate/different simultaneous users won't use a
> common connection so for this very specific scenario keep-alive won't
> matter. Of course for every individual user keep-alive will matter but this
> aspect for the moment I won't to ignore in testing.

It does matter, as there are no clients that request 1KB files all day. You
will see a limited number of full TLS handshakes per second in reality,
because there is keep-alive and there is TLS resumption.

Of course every single user causes at the very least 1 TLS handshake.
That's why ab would have to be *properly* configured to reflect that.

> For this specific scenario (separate/different simultaneous users) is this
> ratio http/https (4600/550) abnormal?
> PS: with Apache/2.4.18 (Ubuntu) the ration is 40% lower

That depends: how many nginx workers do you have compared to
how many apache threads and how does your per-core CPU load
look like when benchmarking?

More information about the nginx mailing list