Skip to content

HTTPS Local Performance

Measurements come from independent local Release loopback processes, not CI.

Metric definitions and methodology.

Average process CPU uses 100% for one logical CPU and can exceed 100% with multiple threads; estimated values are labeled, and incomplete sets of three CPU samples show N/A. On this 14-logical-CPU host, machine share = process CPU% / 14; 100% of one CPU is 7.14% of the machine.

No added computation

Throughput

p99 RTT

All Measured Profiles

64 B / 1 clients / window=1

Dependency / model Round trips/s MiB/s p50 / us p95 / us p99 / us CPU / % RSS / MiB
Boost.Asio / 2 contexts / 2 threads 20,294 [20,113, 20,316] 2.48 47.88 53.92 58.96 53.7 (estimated) 14.25
Boost.Asio / 4 contexts / 4 threads 20,356 [20,315, 20,390] 2.48 47.88 53.75 59.58 53.5 (estimated) 14.36
Boost.Asio / 1 context / 1 thread 19,858 [19,554, 19,901] 2.42 49.17 56.12 63.58 45.8 (estimated) 14.22
Boost.Asio / 1 context / 2 threads 30,134 [29,840, 30,171] 3.68 32.96 36.88 41.92 90.2 (estimated) 14.27
Boost.Asio / 1 context / 4 threads 25,721 [24,572, 26,175] 3.14 37.96 44.96 58.21 144.6 (estimated) 14.41
Standalone Asio / 2 contexts / 2 threads 20,331 [20,112, 20,402] 2.48 47.88 54.38 61.46 53.4 (estimated) 14.53
Standalone Asio / 4 contexts / 4 threads 20,451 [20,405, 20,462] 2.50 47.88 53.62 58.71 53.2 (estimated) 14.59
Standalone Asio / 1 context / 1 thread 19,934 [19,878, 20,000] 2.43 49.17 55.12 61.75 45.7 (estimated) 14.50
Standalone Asio / 1 context / 2 threads 30,231 [29,865, 30,398] 3.69 33.00 37.00 41.12 89.7 (estimated) 14.47
Standalone Asio / 1 context / 4 threads 26,312 [25,915, 26,496] 3.21 37.25 43.79 48.62 144.5 (estimated) 14.64

1024 B / 1 clients / window=1

Dependency / model Round trips/s MiB/s p50 / us p95 / us p99 / us CPU / % RSS / MiB
Boost.Asio / 2 contexts / 2 threads 13,002 [13,001, 13,031] 25.39 75.33 83.33 90.75 41.1 (estimated) 14.22
Boost.Asio / 4 contexts / 4 threads 12,966 [12,957, 12,977] 25.32 75.38 83.38 89.88 41.2 (estimated) 14.41
Boost.Asio / 1 context / 1 thread 12,861 [12,852, 12,867] 25.12 75.38 84.04 90.88 36.1 (estimated) 14.27
Boost.Asio / 1 context / 2 threads 27,126 [27,074, 27,234] 52.98 35.75 40.58 44.17 92.2 (estimated) 14.23
Boost.Asio / 1 context / 4 threads 22,481 [22,271, 22,671] 43.91 43.29 47.38 53.25 132.9 (estimated) 14.38
Standalone Asio / 2 contexts / 2 threads 12,942 [12,935, 12,972] 25.28 75.33 83.71 90.79 41.0 (estimated) 14.41
Standalone Asio / 4 contexts / 4 threads 12,940 [12,915, 13,003] 25.27 75.38 84.00 90.83 41.4 (estimated) 14.64
Standalone Asio / 1 context / 1 thread 12,851 [12,840, 12,872] 25.10 75.38 84.25 90.67 36.1 (estimated) 14.44
Standalone Asio / 1 context / 2 threads 27,256 [26,993, 27,431] 53.23 35.25 40.58 44.38 92.5 (estimated) 14.62
Standalone Asio / 1 context / 4 threads 22,813 [22,478, 22,890] 44.56 43.04 47.08 53.33 133.5 (estimated) 14.61

16384 B / 1 clients / window=1

Dependency / model Round trips/s MiB/s p50 / us p95 / us p99 / us CPU / % RSS / MiB
Boost.Asio / 2 contexts / 2 threads 2,051 [2,042, 2,065] 64.08 467.83 524.92 558.67 27.9 (estimated) 14.19
Boost.Asio / 4 contexts / 4 threads 2,036 [2,010, 2,056] 63.61 470.08 533.00 578.04 28.3 (estimated) 14.31
Boost.Asio / 1 context / 1 thread 2,079 [2,063, 2,079] 64.98 460.00 530.67 551.21 25.9 (estimated) 14.08
Boost.Asio / 1 context / 2 threads 13,900 [13,809, 13,908] 434.39 62.50 67.54 72.79 112.6 (estimated) 14.38
Boost.Asio / 1 context / 4 threads 12,741 [12,644, 12,763] 398.14 67.79 76.21 83.58 126.0 (estimated) 14.45
Standalone Asio / 2 contexts / 2 threads 2,037 [2,022, 2,049] 63.66 470.50 537.92 564.58 28.4 (estimated) 14.55
Standalone Asio / 4 contexts / 4 threads 2,040 [2,035, 2,047] 63.77 470.29 535.21 562.83 28.3 (estimated) 14.55
Standalone Asio / 1 context / 1 thread 2,051 [2,017, 2,062] 64.10 463.83 540.67 569.38 26.6 (estimated) 14.44
Standalone Asio / 1 context / 2 threads 14,051 [13,966, 14,194] 439.08 61.58 67.33 74.42 112.7 (estimated) 14.66
Standalone Asio / 1 context / 4 threads 12,785 [12,688, 12,819] 399.54 67.29 76.29 83.33 126.1 (estimated) 14.66

64 B / 16 clients / window=16

Dependency / model Round trips/s MiB/s p50 / us p95 / us p99 / us CPU / % RSS / MiB
Boost.Asio / 2 contexts / 2 threads 175,116 [173,850, 176,065] 21.38 1,433.96 1,571.38 2,420.17 198.9 (estimated) 34.88
Boost.Asio / 4 contexts / 4 threads 242,978 [239,507, 243,970] 29.66 1,007.33 1,337.83 1,426.38 395.6 (estimated) 35.33
Boost.Asio / 1 context / 1 thread 107,049 [105,963, 108,285] 13.07 2,373.75 2,459.62 2,509.62 99.6 (estimated) 34.69
Boost.Asio / 1 context / 2 threads 152,177 [151,802, 154,215] 18.58 1,659.71 1,766.46 1,947.75 196.0 (estimated) 34.83
Boost.Asio / 1 context / 4 threads 184,173 [182,491, 188,316] 22.48 1,352.79 1,555.92 1,800.50 374.9 (estimated) 35.05
Standalone Asio / 2 contexts / 2 threads 179,694 [175,930, 182,835] 21.94 1,382.67 1,610.00 2,586.92 199.6 (estimated) 35.19
Standalone Asio / 4 contexts / 4 threads 249,736 [248,220, 250,321] 30.49 971.46 1,336.50 1,373.88 395.8 (estimated) 35.66
Standalone Asio / 1 context / 1 thread 109,102 [108,352, 110,028] 13.32 2,328.83 2,418.62 2,498.12 99.6 (estimated) 35.08
Standalone Asio / 1 context / 2 threads 156,113 [154,563, 157,233] 19.06 1,624.21 1,766.88 2,050.00 197.1 (estimated) 35.12
Standalone Asio / 1 context / 4 threads 187,737 [178,601, 189,684] 22.92 1,323.42 1,601.04 1,854.12 375.9 (estimated) 35.56

1024 B / 16 clients / window=16

Dependency / model Round trips/s MiB/s p50 / us p95 / us p99 / us CPU / % RSS / MiB
Boost.Asio / 2 contexts / 2 threads 137,989 [136,530, 142,428] 269.51 1,827.12 1,961.12 2,211.96 196.5 (estimated) 36.19
Boost.Asio / 4 contexts / 4 threads 200,526 [199,385, 203,190] 391.65 1,241.00 1,513.29 1,704.33 378.0 (estimated) 36.66
Boost.Asio / 1 context / 1 thread 70,758 [70,632, 71,113] 138.20 3,606.08 3,697.50 3,838.04 91.5 (estimated) 36.02
Boost.Asio / 1 context / 2 threads 120,762 [119,913, 121,133] 235.86 2,105.79 2,192.29 2,494.25 196.3 (estimated) 36.36
Boost.Asio / 1 context / 4 threads 148,840 [143,913, 151,042] 290.70 1,680.92 1,983.29 2,245.42 377.2 (estimated) 36.34
Standalone Asio / 2 contexts / 2 threads 139,721 [134,804, 142,361] 272.89 1,814.17 1,935.79 2,035.75 195.5 (estimated) 36.30
Standalone Asio / 4 contexts / 4 threads 201,031 [199,232, 201,977] 392.64 1,245.04 1,520.38 1,705.38 373.9 (estimated) 36.86
Standalone Asio / 1 context / 1 thread 76,411 [72,148, 77,229] 149.24 3,359.33 3,457.71 3,561.71 97.0 (estimated) 36.22
Standalone Asio / 1 context / 2 threads 122,440 [121,199, 123,559] 239.14 2,077.42 2,155.46 2,282.58 197.5 (estimated) 36.27
Standalone Asio / 1 context / 4 threads 153,770 [152,089, 154,198] 300.33 1,639.12 1,826.88 2,188.08 380.9 (estimated) 36.34

16384 B / 16 clients / window=16

Dependency / model Round trips/s MiB/s p50 / us p95 / us p99 / us CPU / % RSS / MiB
Boost.Asio / 2 contexts / 2 threads 27,748 [27,630, 27,897] 867.13 9,199.00 9,406.08 9,819.88 165.1 (estimated) 47.66
Boost.Asio / 4 contexts / 4 threads 36,513 [36,076, 36,613] 1,141.02 6,970.50 7,404.04 7,572.17 286.2 (estimated) 47.73
Boost.Asio / 1 context / 1 thread 13,860 [13,754, 13,929] 433.14 18,386.62 18,619.67 18,986.46 86.7 (estimated) 47.38
Boost.Asio / 1 context / 2 threads 30,848 [29,985, 30,864] 963.99 8,289.00 8,526.25 8,841.92 197.7 (estimated) 47.66
Boost.Asio / 1 context / 4 threads 39,174 [39,129, 39,241] 1,224.18 6,505.17 6,759.83 7,173.88 380.9 (estimated) 47.77
Standalone Asio / 2 contexts / 2 threads 27,761 [27,646, 27,908] 867.55 9,191.33 9,430.83 9,895.75 162.6 (estimated) 47.73
Standalone Asio / 4 contexts / 4 threads 36,285 [36,266, 37,201] 1,133.90 6,999.88 7,325.46 7,613.58 288.5 (estimated) 47.92
Standalone Asio / 1 context / 1 thread 14,047 [13,997, 14,083] 438.97 18,163.17 18,661.54 18,827.46 86.7 (estimated) 47.72
Standalone Asio / 1 context / 2 threads 31,183 [31,128, 31,232] 974.46 8,195.12 8,399.96 8,680.50 197.7 (estimated) 47.69
Standalone Asio / 1 context / 4 threads 38,219 [38,210, 40,229] 1,194.35 6,418.29 7,187.62 9,227.62 380.7 (estimated) 47.83

Tables show repeated-run medians; throughput brackets show min/max, not confidence intervals. Latencies are medians of per-run sampled percentiles, without pooling samples.

Navigation

Type to search…

↑↓ navigate↵ selectEsc close