Latency Benchmark

Setup

This benchmark measures round-trip latency of array multiplication across three frameworks: Rosia, ROS 2 (rclpy), and Ray.

Task

A sender node sends a float64 numpy array and a scalar multiplier to a multiplier node, which multiplies the array element-wise and sends the result back. The latency is measured as the time from sending the array to receiving the result.

Pipeline

Sender ──(array, scalar)──► Multiplier ──(result)──► Sender

• Rosia: Two nodes connected via input/output ports in a feedback loop.
• ROS 2: Two nodes communicating via pub/sub topics (array_data and array_result).
• Ray: A remote actor with a multiply() method called via ray.get().

Environment

• Machine: MacBook Pro with Apple M4 Pro chip
• Container: Docker (ARM64), based on ros:humble

Parameters

• Array sizes: 2 to 1,048,576 elements (2^1 to 2^20), corresponding to 16 bytes to 8 MB of float64 data.
• Warmup: 20 iterations per array size (discarded).
• Measured iterations: 20 per array size.
• Metric: Median latency in milliseconds.

The % faster column shows how much less time Rosia takes compared to ROS 2, calculated as (ROS2_median - Rosia_median) / ROS2_median * 100. For example, if ROS 2 takes 10ms and Rosia takes 2ms, the result is +80%, meaning Rosia completes the same task in 80% less time.

Results

On average, Rosia is 64.9% faster than ROS 2.