CPU scaling benchmark

W=0 W=1 W=2 W=3 W=4 W=6 W=8 W=9 W=10 W=12 W=16 iters: 100M 500M 1B 2B
workers
12 +1 main
iters total
500M
38461538/stream
elapsed
1166.51 ms
total CPU used
13678.85 ms
speedup
11.73×
vs serial
efficiency
90.2%
of 13× ideal
stream spawn ms spawned@ work start@ work end@ work ms reap wait ms
0 (main) 0 35.21 35.21 1021.36 986.15 0
1 1.843 1.85 35.42 1132.94 1097.52 114.27
2 1.568 3.44 26.71 1136.89 1110.18 115.68
3 1.532 4.98 28.36 1118.69 1090.33 100.33
4 1.538 6.53 29.35 1091.11 1061.76 77.21
5 1.575 8.12 35.32 1140.85 1105.53 120.03
6 1.659 9.79 41.19 1163.33 1122.14 142.14
7 1.574 11.37 64.74 1043.32 978.58 38.7
8 1.567 12.95 62.33 1057.31 994.98 38.76
9 3.652 16.61 49.47 1107.55 1058.08 86.37
10 1.453 18.07 82.41 1145.14 1062.73 123.93
11 15.127 33.21 66.12 1062.17 996.05 51.09
12 1.974 35.2 82.39 1097.21 1014.82 80.95
main
w1
w2
w3
w4
w5
w6
w7
w8
w9
w10
w11
w12
    fork+handshake      CPU work      parent reap wait
what this measures
Each stream runs a tight integer LCG loop — working set is one CPU register, no memory access, no shared data. Speedup = sum(stream CPU time) / wall-clock elapsed. Efficiency = speedup / (workers+1). 100% efficiency means perfect linear scaling; less than 100% is the cost of serial fork setup, reap tail, SMT/core contention.