Armeria gRPC performance analysis

I noticed some reasonably large differences in performance between armeria-grpc and upstream grpc on the same business logic and did some profiling - want to capture them and maybe start an investigation. My initial suspicion is flushing on every write?

I’ve attached VisualVm profiles, and here’s a screenshot of hot spots in the armeria version. I don’t believe writevAddress, readAddress, invokeFlush0 are supposed to be so high here.

visualvm-profiles.tar.gz

[ikhoon]

Oops, I clicked down(👎) by mistake. 😅

[anuraaga]

Mentioned it in #2295, but one aspect of having competitive gRPC performance will be to optimize as much of the functionality we have that is not used by RPC frameworks, for example routing.

One idea I haven’t explored much is if it’s possible to have an especially optimized code path when we know the server does not support complex routing. For example, it should be possible for this object to have less memory footprint when not using all the different features

https://github.com/line/armeria/blob/master/core/src/main/java/com/linecorp/armeria/server/DefaultRoutingContext.java#L52

Just a brainstorm 😃 /cc @trustin @minwoox @ikhoon