Also, why is the OkHttp engine so slow? 2000k req...

# ktor

[Marc Knaup]

Also, why is the OkHttp engine so slow?

2000k requests (max 200 in parallel)
Apache: ~9s (standard config)
Jetty: ~12s (standard config)
OkHttp: ~15-36s (1k clients, 1k threads)

I’d like to use Apache but it’s unusable (https://youtrack.jetbrains.com/issue/KTOR-717)
Jetty also unstable (https://youtrack.jetbrains.com/issue/KTOR-1120)
CIO also unstable (https://youtrack.jetbrains.com/issue/KTOR-1126)
OkHttp is sloooow… (https://youtrack.jetbrains.com/issue/KTOR-1121)

[mbonnin]

Maybe there's some headers with okhttp that the server takes more time to process? Can you do the same benchmark with MockHttpServer to take the server out of the equation?

[Marc Knaup]

Weird, the Apache engine uses HTTP/1 while OkHttp uses HTTP/2.
a) Why is /2 slower than /1? Or is it the different engine? Will test.
b) Why does Apache not use /2?

[Marc Knaup]

b) Looks like a documentation bug. Documentation states that it supports HTTP/2 but it doesn’t,

[Marc Knaup]

Added for documentation bug: https://youtrack.jetbrains.com/issue/KTOR-1123

[Marc Knaup]

Still a major difference between the engines when both use HTTP/1.

[gildor]

Apache and jetty based on NIO, Okhttp is blocking, so probably this is why

[Marc Knaup]

Maybe, yes.
I’ve also found out that OkHttp uses GZIP even if I don’t install ContentEncoding in the Ktor client.

[Marc Knaup]

Sill, a 500% difference is quite hefty. There are enough threads.

[Marc Knaup]

I can’t really pick an engine for a high volume of calls 😅

Apache doesn’t support HTTP/2 and freezes randomly.
Jetty is full of race conditions.
OkHttp has major performance issues.

I guess there’s no way around HTTP/1 for now 😞

[Marc Knaup]

CIO is as slow as OkHttp Jetty.
In tests with localhost CIO is 300% slower than Apache.

[gildor]

@Marc Knaup have you tried to set Accept-Encoding: identity for okhttp requests?

[Marc Knaup]

I’ve set it to an empty header to disable compression. It’s no longer requesting gzip but the result is the same.

[Marc Knaup]

Great, CIO also randomly gets stuck completely.
Conclusion is that all four engines have severe problems under high load 😞

[e5l]

@Marc Knaup, the CIO problem is fixed and will be released in 1.4.2

[e5l]

Could you share the benchmark code? Right know we have automation to monitor performance numbers for server. I'm going to make the same for client

[Marc Knaup]

Awesome, looking forward to it 🙏

https://gist.github.com/fluidsonic/d4a370bf742f8f9b801512d8777f533c