16 clients; 64/1024/16384-byte messages; windows 1/4/16/64; three repetitions per configuration.
Both executions use native Asio and fixed-size TCP batches: the server reads a complete batch, applies the same computation to each message, then writes the entire reply. Window 1 records message RTT; larger windows record batch RTT. The public arknet coroutine endpoint API remains TODO.
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 extra application computation
- Standalone Asio: Matched coroutine/callback throughput medians range from 0.92× to 1.08×.
- Standalone Asio: Matched coroutine/callback CPU medians range from 0.98× to 1.02×; below 1 means lower average process CPU. Compare with the throughput ratio.
- Boost.Asio: Matched coroutine/callback throughput medians range from 0.92× to 1.11×.
- Boost.Asio: Matched coroutine/callback CPU medians range from 0.92× to 1.05×; below 1 means lower average process CPU. Compare with the throughput ratio.
10,000 deterministic iterations per echo
- Standalone Asio: Matched coroutine/callback throughput medians range from 0.95× to 1.01×.
- Standalone Asio: Matched coroutine/callback CPU medians range from 0.96× to 1.01×; below 1 means lower average process CPU. Compare with the throughput ratio.
- Boost.Asio: Matched coroutine/callback throughput medians range from 0.93× to 1.02×.
- Boost.Asio: Matched coroutine/callback CPU medians range from 0.95× to 1.01×; below 1 means lower average process CPU. Compare with the throughput ratio.
Complete numeric JSON (p95, run ranges and failures)
Failed Profiles
No failed profiles for this protocol.
If any repetition fails, all three measurements for that profile are excluded from rate, latency and matched-ratio summaries; raw failures are retained.
Throughput and CPU ratios pair repetitions at the same dependency combination, payload, batch size, thread model and computation before taking the median; 1 means equal values. Lower CPU utilization does not necessarily complete more requests; compare throughput too. If any CPU sample is missing, its CPU summary and matched ratio are N/A while throughput may remain valid. These results do not predict arknet callback endpoint versus future coroutine API speed.