Benchmark single-threaded, non-blocking TCP performance

[brson]

After #6165, make some benchmarks to establish a networking performance baseline as we continue building out the parallel task scheduler.

[brson]

Some benchmarks above, but need to do more. Not a blocker for 0.7

[msullivan]

No progress on this, I assume? What is the best way to benchmark network performance for a library?

[alexcrichton]

Today I toyed around a bit, and I just wrote some simple programs that was just a server sending a client a page of 0s at a time (continually). Rust averaged ~750MB/s while Go averaged ~530MB/s. I didn't check much else, but @thestinger got some numbers that indicated that C was almost 2x faster than Rust.

[thestinger]

The performance gap doesn't seem incredibly important since it's very unrealistic that you'll actually need >2GiB/s bandwidth from a single thread. It was spending >70% of the time in an internal kernel copying routine at that point.

[thestinger]

Rust doesn't have any truly non-blocking / AIO APIs, and #17325 means that something else needs to be written before there's anything to benchmark in the standard libraries.