Open hishnash opened 3 years ago
Are you testing this is release mode or debug mode?
I have tested debug and profile builds ( think profile build are almost identical to release right?) I will test again when I get back from the shops.
Looking at the code it does look like it is copying data for every frame, since the default frame size is quite small this does mean a lot! Of copy operations.
@0xTim i can confirm this happens with release builds as well.
Describe the bug
When running the client (or server) and the running app receives a Websocket message that is split over many frames CPU load is massive and it takes a very long time to parse the incoming frames concatenating the buffer and contract the message.
To Reproduce
1) create a simple Vapor app with a websocket server 2) create a WebSocket client (using this branch to that is can send large amounts of data https://github.com/vapor/websocket-kit/pull/96) 3) send a large amount of data to the server (a few MB is enough) 4) Notice how the CPU load on the server spikes and it takes ages to handle the incoming data even through the client was able to send it all very quickly without any real CPU load.
Expected behavior
The server should be able to handle a multi frame ws message without completely locking up the cpu core at 100%.
Environment
MacOS (this happens with or without the https://github.com/vapor/websocket-kit/pull/95)
Additional context
Running a profile suggesss that 71% of the cpu time is spent moving memory around.
and 28% is spend relocating the buffer
This seems related to the
WebSocketFrameSequence.append
method. It might well be better to make theWebSocketFrameSequence
is it possible to create a ByteBuffer like object that just points to the underlying existing buffers rather than copies and decloates them?