When I play with apache bench (ab) and I use the -k flag the output for the runs have this
Keep-Alive requests: 0
However if I try Cowboy for example it seems to report correctly
Keep-Alive requests: 15000
as well as being a bit faster (due to reused connections no doubt). I tried this with a few other languages, golang etc and they seemed to handled keep alives with ab.
Now I did an initial search through Bandit docs but I’m not seeing any settings pertaining to this. I did see a small blurb saying it was supported in http/1.1. Ok hmmmmmmmmmm.
What exactly is ab pushing out? Let’s spin up tcdump
huh http/1.0 so I’m guessing Bandit just doesn’t support keep alives in that version. I don’t know exactly what the difference is, I should probably read the http spec at some point, but I am asking if there’s an easy way I can configure Bandit to support that. No problem if there isn’t an easy way. I’m just having fun. But it does look like other servers do support this…probably for ab use cases
I realize this isn’t a direct answer to your question but Elixir folks have generally preferred to use wrk or wrk2 for HTTP benchmarking. It offers more options, stats, and as I recall handles your type of issue more consistently.
A quick google suggests keep-alive wasn’t officially supported before http 1.1, but was added as a response to people using it with http 1.0 without a spec for it existing. So it’s a question of how spec compliant bandit wants to be. A toggle might be a reasonable middleground.
yeah definitely looks like it wasn’t in the original spec.
I will say looking through some of the bandit source code, it’s really not bad reading at all! I might just mess with in on a branch etc. I have no idea how many people actually use http/1.0 in real life, probably not many…
I’ll echo what others have said here; ab is in general a pretty outdated program that does a lot of oddball stuff like this (IIRC, this was actually strictly out of spec prior to the RFC911x series).