Memory Allocator: Switch from talc to galloc

[jounathaen]

I found major performance issues with the current talc allocator. This one replaces it with galloc.

Just like #1935, this PR is for now only intended for testing purposes. Finding the right allocator requires deeper investigations and benchmarking.

[github-actions]

Benchmark Results

Benchmark	Current: ec4f7e8	Previous: 40e0c6e	Performance Ratio
startup_benchmark Build Time	136.18 s	136.14 s	1.00
startup_benchmark File Size	0.91 MB	0.90 MB	1.01
Startup Time - 1 core	0.93 s (±0.01 s)	0.94 s (±0.02 s)	1.00
Startup Time - 2 cores	0.93 s (±0.01 s)	0.92 s (±0.03 s)	1.01
Startup Time - 4 cores	0.95 s (±0.02 s)	0.96 s (±0.03 s)	0.99
multithreaded_benchmark Build Time	139.93 s	140.97 s	0.99
multithreaded_benchmark File Size	0.96 MB	1.01 MB	0.95
Multithreaded Pi Efficiency - 2 Threads	2.93 % (±14.05 %)	2.17 % (±10.40 %)	1.35
Multithreaded Pi Efficiency - 4 Threads	1.56 % (±7.50 %)	1.51 % (±7.26 %)	1.03
Multithreaded Pi Efficiency - 8 Threads	0.75 % (±3.58 %)	0.77 % (±3.68 %)	0.97
micro_benchmarks Build Time	173.85 s	171.42 s	1.01
micro_benchmarks File Size	0.96 MB	1.01 MB	0.95
Scheduling time - 1 thread	3.54 ticks (±16.97 ticks)	2.77 ticks (±13.29 ticks)	1.28
Scheduling time - 2 threads	1.80 ticks (±8.62 ticks)	1.75 ticks (±8.39 ticks)	1.03
Micro - Time for syscall (getpid)	0.21 ticks (±1.00 ticks)	0.12 ticks (±0.58 ticks)	1.72
Memcpy speed - (built_in) block size 4096	1795.98 MByte/s (±8620.69 MByte/s)	1816.86 MByte/s (±8720.93 MByte/s)	0.99
Memcpy speed - (built_in) block size 1048576	544.58 MByte/s (±2613.99 MByte/s)	745.11 MByte/s (±3576.55 MByte/s)	0.73
Memcpy speed - (built_in) block size 16777216	212.16 MByte/s (±1018.35 MByte/s)	219.45 MByte/s (±1053.36 MByte/s)	0.97
Memset speed - (built_in) block size 4096	1518.99 MByte/s (±7291.14 MByte/s)	1875.00 MByte/s (±9000.00 MByte/s)	0.81
Memset speed - (built_in) block size 1048576	1358.14 MByte/s (±6519.08 MByte/s)	1029.20 MByte/s (±4940.18 MByte/s)	1.32
Memset speed - (built_in) block size 16777216	901.81 MByte/s (±4328.69 MByte/s)	924.64 MByte/s (±4438.25 MByte/s)	0.98
Memcpy speed - (rust) block size 4096	1395.35 MByte/s (±6697.67 MByte/s)	1411.76 MByte/s (±6776.47 MByte/s)	0.99
Memcpy speed - (rust) block size 1048576	554.89 MByte/s (±2663.49 MByte/s)	693.90 MByte/s (±3330.73 MByte/s)	0.80
Memcpy speed - (rust) block size 16777216	209.75 MByte/s (±1006.78 MByte/s)	219.67 MByte/s (±1054.40 MByte/s)	0.95
Memset speed - (rust) block size 4096	1304.35 MByte/s (±6260.87 MByte/s)	1791.04 MByte/s (±8597.01 MByte/s)	0.73
Memset speed - (rust) block size 1048576	1332.53 MByte/s (±6396.16 MByte/s)	1105.00 MByte/s (±5304.01 MByte/s)	1.21
Memset speed - (rust) block size 16777216	889.82 MByte/s (±4271.14 MByte/s)	954.96 MByte/s (±4583.81 MByte/s)	0.93
alloc_benchmarks Build Time	160.26 s	157.27 s	1.02
alloc_benchmarks File Size	0.97 MB	0.97 MB	1.01
Allocations - Allocation success	2.00 % (±13.86 %)	2.00 % (±13.86 %)	1
Allocations - Deallocation success	1.39 % (±9.67 %)	1.40 % (±9.67 %)	1.00
Allocations - Pre-fail Allocations	2.00 % (±13.86 %)	2.00 % (±13.86 %)	1
Allocations - Average Allocation time	708.39 Ticks (±4908.88 Ticks)	262.62 Ticks (±1819.84 Ticks)	2.70
Allocations - Average Allocation time (no fail)	708.39 Ticks (±4908.88 Ticks)	262.62 Ticks (±1819.84 Ticks)	2.70
Allocations - Average Deallocation time	28.48 Ticks (±197.35 Ticks)	17.05 Ticks (±118.18 Ticks)	1.67
mutex_benchmark Build Time	157.81 s	159.65 s	0.99
mutex_benchmark File Size	0.97 MB	1.01 MB	0.95
Mutex Stress Test Average Time per Iteration - 1 Threads	0.30 ns (±2.08 ns)	0.36 ns (±2.49 ns)	0.83
Mutex Stress Test Average Time per Iteration - 2 Threads	0.34 ns (±2.36 ns)	0.38 ns (±2.63 ns)	0.89

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

[m-mueller678]

I am happy that we are getting a faster allocator, but I think galloc is still not particularly good. The benchmark from #1984 reports around 1s for galloc, so still a 20x slow-down compared to linux. I think using a single-threaded allocator and wrapping it in a mutex will never get good performance in a multithreaded setting. Unfortunately, there do not seem to be any allocators with a better synchronization scheme that are also easily portable to hermit.