Skip to content
arknet

Local Performance Report

View as Markdown

The report contains the final measurements for each protocol: UDP on 2026-10-12, and other protocols and native coroutine TCP on 2026-10-10. Each protocol retains its own source hash and raw records.

1,620 local IPv4 measurements compare throughput, p99 latency and IO models across the Asio dependency combinations.

Test Conditions

Configuration Test machine
Machine model Mac16,11
Processor Apple M4 Pro
Architecture arm64
Physical / logical CPUs 14 / 14
Memory 48 GiB
Operating system macOS 27.0.1 (26A434)
Kernel release 27.0.0

Release build; measurement=1.0 s; warmup=0.25 s; repetitions=3.

Main charts: 1024 bytes, 16 clients, no added computation. Inflight messages per connection: UDP 1; other protocols 16. Both peers share the process and IO workers over local loopback.

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.

Key Results

  • TCP / Standalone Asio / No added computation: 4 contexts / 4 threads has this group’s highest throughput median (1.98× one thread); its p99 is 1,346.00 us versus 1,491.79 us with one thread.
  • UDP / Standalone Asio / No added computation: 1 context / 1 thread has this group’s highest throughput median (92,515 round trips/s), with p99 RTT of 207.79 us and process CPU utilization of 99.4%.
  • Coroutine TCP / Standalone Asio / No added computation: 1 context / 4 threads has this group’s highest throughput median (1.84× one thread); its p99 is 870.46 us versus 1,028.71 us with one thread.
  • Coroutine TCP / Standalone Asio / 10,000 CPU iterations per echo: 4 contexts / 4 threads has this group’s highest throughput median (3.54× one thread); its p99 is 1,749.29 us versus 5,687.25 us with one thread.

The highest median applies only to the stated load; complete results appear below.

Throughput

Local throughput overview

p99 Round-Trip Latency

99% of sampled round trips finish within this time; 1000 microseconds = 1 millisecond.

Local p99 latency overview

Choosing an IO Model

Model Workload Tradeoff
1 context / 1 thread Low concurrency, control services, light state handling Long handlers delay every connection
1 context / several threads Uneven connection activity, shared available workers Use per-connection strands; shared scheduling has overhead
Several contexts / 1 thread each Partitioned sessions, tenants or stable assignment A busy shard cannot borrow other shards’ workers

Coroutine IO and CPU workload comparisons.

Detailed Results by Protocol

Detailed pages retain all 64/1024/16384-byte and 1/16-client profiles. Bars show repeated-run medians; error bars show min/max, not confidence intervals.

All measured values (JSON) | Payload, batch and execution comparisons

Coroutine window=16 measures batch RTT; the callback program measures message RTT. These results do not establish coroutine speedups.

Toolchain and Build

Apple Clang 21.0.0; CMake 4.3.3; Ninja; Release -O3 -DNDEBUG -std=gnu++20 -arch arm64; TLS ON; sanitizers OFF. Standalone Asio 1.38.2 with BHO Beast based on Boost 1.84 (API 351); Boost 1.90.0 with Asio 1.38.0 and Beast API 359; OpenSSL 3.6.4. Mac mini on AC power, low power mode disabled, no CPU affinity.

The dependency combinations use different Beast versions, so differences between them also include dependency-version effects and cannot be attributed solely to Asio.

Metric Definitions

Throughput, RTT, CPU, RSS and error definitions.

Raw Data and Reproduction

Scope

Local loopback results do not establish cross-host or production capacity. Short runs, scheduling and thermals introduce variance; repeat with the deployment workload.

Navigation

Type to search…

↑↓ navigate↵ selectEsc close