Add more file reading benchmarks

[threecgreen]

Motivation

I wanted to add more representative benchmarks to compare blocking reads against io_uring for the current AsyncRead and fs::read implementations.

Solution

This PR adds file reading benchmarks that read from actual files of different sizes. The existing fs benchmarks read from /dev/zero. I also added benchmarks for concurrent reading with a small blocking thread pool to try and test the effect of contention.

Results

io_uring performs well on the current_thread runtime, about 30-60% faster than spawn_blocking. I was surprised io_uring performs worse on multi_thread, even on concurrent reads.

Median results (lower is better).

System: AMD Ryzen 9 5950X, 64 GB RAM, Linux 6.19.11
Storage: Samsung SSD 980 NVMe 1TB (ext4)

Sequential fs::read()

size	sync	spawn_blocking	io-uring	vs. sync	vs. spawn_blocking
64KiB	7 us	15 us	38 us	+461%	+147%
1MiB	43 us	55 us	79 us	+84%	+43%
16MiB	1.5 ms	1.9 ms	1.4 ms	-8%	-23%
256MiB	172 ms	177 ms	186 ms	+8%	+5%

Sequential File stream (64K buf, multi_thread)

size	sync	spawn_blocking	io-uring	vs. sync	vs. spawn_blocking
64KiB	6 us	42 us	47 us	+624%	+10%
1MiB	48 us	235 us	323 us	+571%	+38%
16MiB	947 us	5.0 ms	6.0 ms	+534%	+19%
256MiB	27.8 ms	73.5 ms	96.8 ms	+248%	+32%

Sequential File stream (64K buf, current_thread)

size	sync	spawn_blocking	io-uring	vs. sync	vs. spawn_blocking
64KiB	6 us	41 us	29 us	+352%	-28%
1MiB	48 us	253 us	100 us	+108%	-60%
16MiB	947 us	4.5 ms	1.8 ms	+94%	-59%
256MiB	27.8 ms	81.8 ms	44.0 ms	+58%	-46%

Concurrent fs::read() (256 MiB files)

tasks	spawn_blocking	io-uring	vs. spawn_blocking
multi_thread/4	216 ms	770 ms	+257%
multi_thread/16	790 ms	3.2 s	+308%
multi_thread/64	6.4 s	12.4 s	+95%
current_thread/4	219 ms	744 ms	+239%
current_thread/16	785 ms	3.0 s	+277%
current_thread/64	5.6 s	11.8 s	+109%

Concurrent File stream (256 MiB files)

tasks	spawn_blocking	io-uring	vs. spawn_blocking
multi_thread/4	151 ms	229 ms	+52%
multi_thread/16	493 ms	823 ms	+67%
multi_thread/64	1.5 s	2.8 s	+86%
limited_blocking/4	249 ms	229 ms	-8%
limited_blocking/16	509 ms	822 ms	+61%
limited_blocking/64	1.2 s	2.9 s	+143%

[threecgreen]

FreeBSD failure looks flaky: No space left on device (28)

[mattiapitossi]

Probably we would need to run a cargo clean, I’ll open a PR. Can you trigger the CI again with an empty commit?