executor: Don't run background tasks if future is ready

[Gelbpunkt]

Often times, the future is ready immediately and yet we will still run the background tasks every time, adding significant overhead that is noticable when benchmarking a HTTP server.

The blocking httpd example gains anywhere between 20-50% more throughput for me with these changes, while the netbench TCP bandwidth benchmark only slightly regresses by about 5%. The httpd example ported to axum sees an increase in throughput by 500-600%:

before:

Beginning round 1...
Benchmarking 128 connections @ http://127.0.0.1:9975 for 15 second(s)
  Latencies:
    Avg      Stdev    Min      Max
    6.27ms   83.92ms  1.25ms   11608.18ms
  Requests:
    Total: 212052  Req/Sec: 14143.12
  Transfer:
    Total: 43.26 MB Transfer Rate: 2.89 MB/Sec

after:

Beginning round 1...
Benchmarking 128 connections @ http://127.0.0.1:9975 for 15 second(s)
  Latencies:
    Avg      Stdev    Min      Max
    0.57ms   0.56ms   0.12ms   209.94ms
  Requests:
    Total: 1197609 Req/Sec: 79857.55
  Transfer:
    Total: 244.29 MB Transfer Rate: 16.29 MB/Sec

This was suggested by @zyuiop in #2086 (comment).

[zyuiop]

Thanks for bringing back this patch!

I'm concerned the fact that the executor may run way less often may cause other issues. Then again, the run call here is only invoked when calling block_on, so if calling run less often causes issues, this means it should probably be invoked in a timer interrupt or something like that.

[mkroening]

I'm concerned the fact that the executor may run way less often may cause other issues. Then again, the run call here is only invoked when calling block_on, so if calling run less often causes issues, this means it should probably be invoked in a timer interrupt or something like that.

Yeah, I think this is happening in CI. Of course, one of the issues with our executor is that many things are actually blocking, which is bad. Making things properly non-blocking should make polling the executor here a non-issue. On the other hand, I agree that if something breaks due to this PR, that is probably a bug and should be resolved.

[github-actions]

Benchmark Results

Details

Benchmark	Current: e67a10f	Previous: 7c500d8	Performance Ratio
startup_benchmark Build Time	90.87 s	91.25 s	1.00 ❗
startup_benchmark File Size	0.81 MB	0.79 MB	1.02 ❗
Startup Time - 1 core	0.96 s (±0.03 s)	0.93 s (±0.03 s)	1.03
Startup Time - 2 cores	0.97 s (±0.03 s)	0.94 s (±0.03 s)	1.03
Startup Time - 4 cores	0.99 s (±0.03 s)	0.94 s (±0.02 s)	1.05 ❗
multithreaded_benchmark Build Time	86.72 s	88.17 s	0.98 ❗
multithreaded_benchmark File Size	0.91 MB	0.89 MB	1.02 ❗
Multithreaded Pi Efficiency - 2 Threads	88.71 % (±7.78 %)	90.13 % (±10.09 %)	0.98
Multithreaded Pi Efficiency - 4 Threads	43.62 % (±3.09 %)	45.06 % (±4.71 %)	0.97
Multithreaded Pi Efficiency - 8 Threads	25.50 % (±1.63 %)	25.74 % (±2.31 %)	0.99
micro_benchmarks Build Time	94.77 s	93.18 s	1.02 ❗
micro_benchmarks File Size	0.92 MB	0.90 MB	1.02 ❗
Scheduling time - 1 thread	72.96 ticks (±3.77 ticks)	71.27 ticks (±4.26 ticks)	1.02
Scheduling time - 2 threads	40.96 ticks (±6.09 ticks)	39.04 ticks (±5.20 ticks)	1.05
Micro - Time for syscall (getpid)	3.09 ticks (±0.25 ticks)	2.97 ticks (±0.30 ticks)	1.04
Memcpy speed - (built_in) block size 4096	65436.08 MByte/s (±47028.44 MByte/s)	66371.85 MByte/s (±47142.10 MByte/s)	0.99
Memcpy speed - (built_in) block size 1048576	29657.45 MByte/s (±24588.04 MByte/s)	29496.16 MByte/s (±24371.57 MByte/s)	1.01
Memcpy speed - (built_in) block size 16777216	27575.23 MByte/s (±23021.49 MByte/s)	27947.10 MByte/s (±23282.80 MByte/s)	0.99
Memset speed - (built_in) block size 4096	66053.38 MByte/s (±47416.08 MByte/s)	66660.22 MByte/s (±47324.52 MByte/s)	0.99
Memset speed - (built_in) block size 1048576	30460.31 MByte/s (±25032.45 MByte/s)	30239.65 MByte/s (±24809.14 MByte/s)	1.01
Memset speed - (built_in) block size 16777216	28336.00 MByte/s (±23452.91 MByte/s)	28701.07 MByte/s (±23711.18 MByte/s)	0.99
Memcpy speed - (rust) block size 4096	56917.41 MByte/s (±42026.53 MByte/s)	59143.47 MByte/s (±43480.35 MByte/s)	0.96
Memcpy speed - (rust) block size 1048576	29396.25 MByte/s (±24539.08 MByte/s)	29331.43 MByte/s (±24300.99 MByte/s)	1.00
Memcpy speed - (rust) block size 16777216	27420.15 MByte/s (±22866.37 MByte/s)	28046.81 MByte/s (±23386.04 MByte/s)	0.98
Memset speed - (rust) block size 4096	57564.84 MByte/s (±42408.57 MByte/s)	59676.52 MByte/s (±43806.15 MByte/s)	0.96
Memset speed - (rust) block size 1048576	30204.88 MByte/s (±24996.22 MByte/s)	30122.22 MByte/s (±24741.83 MByte/s)	1.00
Memset speed - (rust) block size 16777216	28032.05 MByte/s (±23163.63 MByte/s)	28769.44 MByte/s (±23781.83 MByte/s)	0.97
alloc_benchmarks Build Time	93.51 s	92.51 s	1.01 ❗
alloc_benchmarks File Size	0.88 MB	0.86 MB	1.02 ❗
Allocations - Allocation success	100.00 %	100.00 %	1
Allocations - Deallocation success	100.00 %	100.00 %	1
Allocations - Pre-fail Allocations	100.00 %	100.00 %	1
Allocations - Average Allocation time	11197.85 Ticks (±194.00 Ticks)	10757.47 Ticks (±159.10 Ticks)	1.04 ❗
Allocations - Average Allocation time (no fail)	11197.85 Ticks (±194.00 Ticks)	10757.47 Ticks (±159.10 Ticks)	1.04 ❗
Allocations - Average Deallocation time	1082.14 Ticks (±540.08 Ticks)	1302.69 Ticks (±790.51 Ticks)	0.83
mutex_benchmark Build Time	92.90 s	93.08 s	1.00 ❗
mutex_benchmark File Size	0.91 MB	0.90 MB	1.02 ❗
Mutex Stress Test Average Time per Iteration - 1 Threads	13.10 ns (±0.70 ns)	13.32 ns (±0.71 ns)	0.98
Mutex Stress Test Average Time per Iteration - 2 Threads	23.40 ns (±14.93 ns)	18.58 ns (±6.34 ns)	1.26

This comment was automatically generated by workflow using github-action-benchmark.