Multiple accept loops in named pipes transport

[JamesNK]

Feedback from #14207

PR adds multiple accept queues. I did some very cursory benchmarks via a unit test and found an improvement when 2 or more queues were used.

Creating 50,000 connections.

• 1 accept queue: 6.1 sec
• 2 accept queues: 4.7 sec

I didn't notice any improvement from increasing the queue count above 2, but in other places in transports, we are using Environment.ProcessorCount, so I followed that.

[JamesNK]

cc @simonferquel

[simonferquel]

Small nits, increasing the bounded channel capacity might be helpful

[mgravell]

two nits; the captured i seems a definite "this should be fixed"

[mgravell]

👍 on the pragmatic solution to those logger bits :)

[JamesNK]

👍 on the pragmatic solution to those logger bits :)

It was just some Console.WriteLine like debugging 😄

[davidfowl]

@JamesNK feels like we should have a benchmark for this in the perf infrastructure.

[JamesNK]

Using a unit test as a benchmark is gross. Named pipes benchmarks would differ from our regular tests because the client and server are on the same machine.

@sebastienros Are there existing examples of benchmarks running the client and server processes on one machine?

[sebastienros]

@JamesNK here is how it's done, same architecture (multiple applications) but you can see where main and application are used.

https://github.com/aspnet/Benchmarks/blob/main/scenarios/redis.benchmarks.yml#L36-L45

And in profiles, we defined main and application to be aliases of the same machine, so crank will send the two jobs to the same machine. Being in the same run the agent won't queue the second one end them both.

https://github.com/aspnet/Benchmarks/blob/main/scenarios/aspnet.profiles.standard.yml#L56-L60

So the only special thing to do is to define the job with main and application as service names.

[JamesNK]

Blocked on API review: #46268

[halter73]

I'm interested in seeing the results on our "Ampere Linux (80)" machine. I know you said that perf didn't improve on your machine adding more than 2 accept loops, but did it get any worse before hitting the limit of 16?