IO Models and Coroutine Performance

Native Asio Coroutine Baseline

Native Asio TCP coroutine Release local loopback: 1024-byte messages, 16 clients and window 16. Latency is batch RTT.

No Extra Application Computation

Coroutine throughput by IO model

Native Asio coroutine batch p99 RTT by IO model

One context/four threads has the highest throughput median and one context/two threads the lowest batch p99 median for both dependency combinations.
Standalone’s batch p99 run ranges overlap between one context/two threads and one context/four threads, so these runs do not establish a consistent tail-latency advantage.

10,000 Deterministic Iterations Per Echo

Native Asio coroutine throughput with application computation

Native Asio coroutine batch p99 RTT with application computation

Four contexts/one thread each has the highest throughput median and lowest batch p99 median for both dependency combinations.
Batch p99 run ranges overlap between one context/four threads and four contexts/one thread each for both dependency combinations.

Complete coroutine baseline · Payload, batch and execution comparisons

Test environment · Metrics and statistics · TCP callback results

Reproducing the Comparison

The checked-in programs benchmarks/loopback.cpp (arknet callbacks) and benchmarks/coroutine.cpp (native Asio callbacks/coroutines) each have an independent main. Build with CMake and run the C++ executable directly:

cmake -S . -B build-perf -DCMAKE_BUILD_TYPE=Release \
  -DARKNET_BUILD_BENCHMARKS=ON -DARKNET_ENABLE_SSL=OFF \
  -DARKNET_USE_BOOST_ASIO=OFF -DARKNET_ASIO_INCLUDE_DIR=/path/to/asio/include
cmake --build build-perf --target arknet_coroutine_benchmark --parallel 3
build-perf/benchmarks/arknet_coroutine_benchmark \
  --execution coroutine --protocol tcp --payload 1024 --clients 16 --window 16 \
  --io-model shared --io-threads 4 --work 0 --warmup 0.25 --seconds 1

This example uses one context/four threads, 1024-byte messages, 16 clients and window 16, writing JSON to standard output. Replace the Asio include argument with -DARKNET_USE_BOOST_ASIO=ON for Boost. With Visual Studio, add --config Release and run the .exe under benchmarks/Release/.

The --io-model values shared and sharded select one context with multiple threads and one context per thread, respectively. Set --io-threads 1 to test one context/one thread. --execution callback runs the matched native callback mode. --work 10000 adds 10,000 deterministic iterations per server echo; 0 adds no application computation. This is a synthetic load, not a production parser or crypto algorithm. Compare topologies within the same benchmark program, dependency combination, payload, concurrency and computation setting. The Asio coroutine program sends and reads fixed-size batches: with --window above one, its latency is batch RTT. The arknet callback program continuously replenishes its window and reports individual-message RTT. Their throughput and latency do not establish a coroutine-versus-callback speedup.

The expanded comparison runs callback and coroutine modes of the same native Asio TCP program. Its server reads a complete batch, performs the configured computation on each message, then writes the full reply. Window 1 measures message RTT; larger windows measure batch RTT. See the matched comparison.

Repeat configurations in fresh C++ processes, or use the maintainer’s shell matrix in Testing. Only that batch tool needs jq, GNU timeout and shasum; direct C++ tests and single benchmarks do not.

Raw results and test commands are documented in Testing and Performance. Short loopback measurements cannot predict WAN latency or production capacity.