Performance

These reports measure local loopback workloads on a Mac mini M4 Pro, not CI runner performance. Choose a report with matching protocol, payload, concurrency and execution mode; each report records its host, source, commands and scope.

Download the raw measurements and verification tools with the SHA256 checksum. The data is distributed as a release asset, outside the source repository.

Version and execution comparisons

• Database clients: MySQL and Redis throughput, response time and CPU for v0.2.0 and the complete-shutdown implementation.
• Network endpoints: matched TCP, TLS TCP, HTTP, HTTPS, WebSocket and WSS workloads.
• Five-second workload comparison: eight selected workloads, five fresh processes per version, with throughput, CPU and response time.
• Asio callbacks and coroutines: execution modes and worker counts within the same revision.

The configuration grids contain 1,632 rounds; the five-second comparison contains 80 rounds. Databases use native connection callbacks; the Asio control bypasses arknet endpoints. They do not measure ark’s database coroutine bridge.

RUDP measurements

RUDP FEC report: 96 configurations and 276 measured runs across both Asio dependency combinations, three IO models, ARQ-only/fixed/adaptive policies, payloads, in-flight requests and injected loss. SVG charts compare throughput, p99 RTT, CPU, wire overhead, recovery, retransmissions and retained memory. The public RUDP API currently uses callbacks; RUDP coroutines remain TODO.

Database clients

• MySQL: read-only text and prepared queries, payloads, connections and client queue depth; query RT, throughput, CPU and memory.
• Redis: binary ECHO with pipeline sizes and in-flight batches; whole-batch RT, command throughput, CPU and memory.

Both reports compare Boost.Asio and Standalone Asio, plain TCP and verified TLS. They use callback clients; connection pools and database coroutines are outside these measurements.

UDP measurements

UDP performance report includes 504 matrix runs, 21 buffer probes and a 60-second stress test for each dependency combination.

• IO models: one context/one thread, one context/multiple threads and partitioned contexts.
• Payload and in-flight requests: repeated configurations with failure groups retained.
• Limits and loss: datagram sizes and receive-queue overflow.

Callback, coroutine and CPU comparisons

Matched native TCP measurements compare callbacks and coroutines under the same batch workload. IO model comparisons explain the original scheduler results and their measurement boundaries. These programs use native Asio; public arknet coroutine endpoints remain TODO.

Protocol results

The multi-protocol summary and dimension summary show the final results for each measured workload, with dates and source hashes. Measurements do not establish WAN capacity, IPv6 performance or production endurance.

Methodology and metric definitions · Correctness validation · Buffer configuration