Async JSON read/write/(de)serialize

[TahirAhmadov]

Does it make sense to make JSON operations async, with ValueTask and writing to a PipeWriter? I'm concerned about potential performance hit of all the allocations of async state machines. Has anybody looked into it at all?

[FiniteReality]

I believe that if a method returns ValueTask, you can expect it to run completely synchronously without allocating a state machine for as long as it can. In this case, it would be specifically when the PipeWriter buffer is empty.

[FiniteReality]

I did, my apologies. Looking at the code, there seems to only be the allocation of a single Task within the state machine, and the implementation avoids further allocations.

[TahirAhmadov]

Yes, but the state machine itself is a heap object which is always allocated. Meaning, with Task, it's actually two allocations per async method call.

[FiniteReality]

It doesn't seem to be a heap object: https://sharplab.io/#v2:EYLgxg9gTgpgtADwGwBYA+ABATARgLABQGADAAQY4CsA3IYRgMzlakDCpA3oaT+UxgA5SANQCGAGwCuMDElIBZABQBKTt14aKATkUAiUbuW0CGzVtIA5FcdO9te4IZu3y5q0fW37usE8+kAX39/RhEJaVlLFTUTFwwAdlIAExgAM1FJcQAXZx4gggCgA

[TahirAhmadov]

Oh yes, you are right - I was looking at IL generated by Task assumed it did a heap alloc with ValueTask, as well.
I guess the remaining question is, does the non-allocation overhead add up to making it more expensive than the benefits of writing JSON async.

[TahirAhmadov]

Clarification: that was old, .NET 4.8 code which allocates a heap object for state machine. In .NET 6, state machines are always a struct, regardless of Task or ValueTask.

[TahirAhmadov]

I ran a quick benchmark and it looks like when writing to a StringBuilder, everything else being equal, (.NET 6) a collection of 10,000 DTOs is JSON-written in: (average of 10 runs)
Sync: 00:00:00.0309208
Async: 00:00:00.0651994

[TahirAhmadov]

To answer my own question: it does not seem to be worth it to make methods of a JsonWriter-like class asynchonous, even with ValueTask, and this includes HtmlTextWriter. I've actually written an async HtmlTextWriter implementation and code to use it, and am going to revert to synchronous. The performance penalty, rendering a moderately sized HTML page, is humongous; something on the order of going from 1ms to about 100ms.