Better defaults for the Pipe

[davidfowl]

• Use a 4K buffer instead of 2K (4K is a very common buffer size and usually maps to system page size)
• Use a stack for the buffer segment pool and allow pooling more than 16 segments for large writes (up to 256 max)

[benaadams]

As an aside:

A pattern useful to maintain to reduce syscalls for Pipe for file reading (that Stream avoids by defaulting to a very large buffer), is using ReadFileScatter on Windows and readv, preadv on Linux to write to multiple segments at once.

This ends up being something like:

GetMemory, Advance, GetMemory, Advance, GetMemory, Advance, GetMemory, ReadGather, FlushAsync

Mentioning it as it doesn't work with the IBufferWriter<T>.Advance advice change dotnet/dotnet-api-docs#1949 (Advance invalidates previous span/memory)

However, its needed to implement IHttpSendFileFeature for Kestrel directly over the Body pipes without increasing the syscall count by about x 20 vs Stream (which has the disadvantage of an 85kB temp buffer alloc and secondary copy).

[JanEggers]

if it should be system page size why not use this as default?
https://docs.microsoft.com/en-us/dotnet/api/system.environment.systempagesize

[benaadams]

if it should be system page size why not use this as default?

If you are using large pages then the page size can be 2MB to 256MB which would be an incredibility inefficient use of memory across multiple pipes.

[pakrym]

What's the reason for using the stack here? It's another allocation per pipe and it's 3 fields instead of 2 that we had before.

[davidfowl]

We can live with another field. Also we saved on one the other PR so it’s a trade

[pakrym]

But what are we gaining by spending a field and an allocation?

[davidfowl]

Clearer code and a way to resize the initial buffer up to the max without writing that code manually.

[pakrym]

It's literally 3 lines of code:

corefx/src/System.Collections/src/System/Collections/Generic/Stack.cs

Lines 286 to 289 in 2a2d34d

	Array.Resize(ref _array, (_array.Length == 0) ? DefaultCapacity : 2 * _array.Length);
	_array[_size] = item;
	_version++;
	_size++;

[pakrym]

How big of an improvement does this give? What scenario is driving this change?

[davidfowl]

Recently testing out benchmarks with JSON tokens bigger than 32K and also the fact that pipelines is exposed in ASP.NET Core means you can buffer more (and without the hard limit). This makes it so that we don’t start allocating segments until you reach a 1MB of buffered data by default.

Also the 2K vs 4K by default felt like a better buffer size (we use it pretty much everywhere else).

[halter73]

Nit: MaxSegmentPoolSize.