Port benchmarks for the websocket encoder/decoder.

[amizurov]

Motivation:

Porting benchmarks for the websocket encoder/decoder.

Modification:

Add WebSocketFrame13DecoderBenchmark and WebSocketFrame13EncoderBenchmark benchmarks.

Result:

We can  compare changes related to websocket performance.

[chrisvest]

Use allocator.constBufferSupplier in the setUp, because after the first call to split, the content buffer will be empty.

[chrisvest]

Same in the decoder benchmark.

[amizurov]

@chrisvest Thanks, got it. Hmm, i also see that EmbeddedEventLoop has changed its behavior. In 4.1 it always returns true when calling inEventLoop(), now only if a task running inside the current loop. I think we need to review all benchmarks because now they may not work correctly.

[amizurov]

@chrisvest PTAL again

[amizurov]

@chrisvest @normanmaurer As i mentioned the EmbededEventLoop has changed its behaviour and i'm not found a better solution than executing benchmark inside an embedded event loop to proper handling writePromiseCombiner(ctx, out, promise). Maybe you can advice something else. Also bellow the benchmark results according changes (#12780) the cost of split() is significantly decrease performance.

[chrisvest]

The build failed because it needs a rebase to fix imports.

I don't see split showing up much in the profile. Rather, it's the always-on cleaners; CleanerDrop.innerWrap accounts for 17.6% of the time in a run I just did. I plan to make them off by default, and that should bring performance much closer.

[amizurov]

The build failed because it needs a rebase to fix imports.

I don't see split showing up much in the profile. Rather, it's the always-on cleaners; CleanerDrop.innerWrap accounts for 17.6% of the time in a run I just did. I plan to make them off by default, and that should bring performance much closer.

Thanks, I'll be happy to double-check after disabling, because it's inconvenient to manually close the incoming buffer in codecs.

[chrisvest]

I played around a bit; the smaller the contentLength, the bigger impact CleanerDrop.innerWrap has. 43.9% with contentLength = 100. The 17% was from contentLength = 3000. Using off-heap allocator and masking = false.

[chrisvest]

Thanks