I ve switched out Netty for CIO as the Ktor server to see if

Question

I ve switched out Netty for CIO as the Ktor server to see if the performance would be better but it seems much worse Is CIO supposed to outperform Netty now or at some point or can it be tweaked to do that On Netty my throughput is about 38 000 requests per minute with 2 5 seconds per request On CIO it s only about 21 000 and each request takes almost 5 seconds

janvladimirmostert · Answer

i see there are things that can be tweaked on CIO are these values the defaults or are there no limits by default <https ktor io clients http client engines html cio>

janvladimirmostert · Answer

seems like those are the defaults doesn t seem like they can be set when using the embeddedServer option <https github com ktorio ktor blob master ktor client ktor client cio jvm src io ktor client engine cio CIOEngineConfig kt>

Jorge R · Answer

< janvladimirmostert> can I ask you how do you get those metrics

Jorge R · Answer

I wont be able to answer your question but I would also like to get that info out of Ktor

janvladimirmostert · Answer

that would be <http loader io|loader io> i created a test that simulates 2000 users concurrently hitting a specific page that i m trying to optimize for spikes in traffic Obviously there s a couple of bottlenecks i should still fix at 500 users the load times are still around 0 5 seconds where i want it as soon as the load goes to 2000 concurrent users that spikes to 2 5 seconds per request which is too slow this was on Netty When i now swop out Netty for CIO Co routine IO and do literally nothing else to the code run a bunch of warm up load tests the latency almost doubles throughput halves I m thinking it has something to do with the request limit per page which seems to be set to 100 not sure how to change that when using the embedded configuration

janvladimirmostert · Answer

``` class EndpointConfig Maximum connections per single route var maxConnectionsPerRoute Int = 100 ``` 100 is seriously low when you re trying to squeeze a lot of performance out of a single instance

janvladimirmostert · Answer

here s an almost copy paste of my server config <https stackoverflow com questions 57978856 respond with cache headers for images in ktor>

janvladimirmostert · Answer

all that changes now is adding the CIO depedency replacing Netty with CIO and adding the experimental annotation on top of main

janvladimirmostert · Answer

``` JvmStatic KtorExperimentalAPI fun main args Array lt String gt ``` ``` embeddedServer Netty CIO watchPaths = listOf module module = Application module port = if ENV env == LOCAL 8080 else 80 apply start wait = true ```

janvladimirmostert · Answer

have not looked very deep into it will play around with it further tonight and see if i can get that 5 seconds down to under 1 second

avolkmann · Answer

Last thing I remember was that CIO performance will be optimized later

janvladimirmostert · Answer

ok any idea if there s a rough ETA for that going to switch back to Netty until then