Performance degradation when adding subscriber over Ethernet

[mariamercade]

Hello,

I’m experiencing communication problems using Cyclone DDS with ROS2 Jazzy. Below are the details:

Setup

Robot CPU: Intel i7, 8GB RAM, Ubuntu 24.04 Server

    Runs all the robot software.

    IP: 192.168.33.198

Development VM: Ubuntu 24.04 Desktop, 8GB RAM, 6 CPUs (host is also an Intel i7 running Windows)

    Used for development and testing.

    IP: 192.168.33.197

Both connected via Ethernet (100 Mbps), same subnet. I want to have smooth topic communication between them through Ethernet.

Tests performed

On the CPU (robot), I have a ROS 2 publisher of a PointCloud2 topic publishing at ~10Hz with Best Effort QoS.

I ran the following tests:

Test 1 — CPU only, with <AllowMulticast>true</AllowMulticast>:

• Publisher and subscribers running on the CPU.
• Adding subscribers caused the topic rate to drop significantly as I added more.

Test 2 — CPU only, with <AllowMulticast>spdp</AllowMulticast>:

• Publisher and subscribers running on the CPU.
• The rate problem was resolved: the publisher maintained ~10Hz even with multiple subscribers.
• (I’m not entirely sure why setting <AllowMulticast>spdp</AllowMulticast> fixed this.)

Test 3 — Publisher on CPU, subscriber on VM, both with <AllowMulticast>spdp</AllowMulticast>:

• Publisher runs on CPU.
• Single subscriber runs on the VM over Ethernet.
• The topic rate drops drastically to about ~5.5Hz — even with just one subscriber on the VM

So, the change to <AllowMulticast>spdp</AllowMulticast> solved the issue on a single machine, but subscribing from the VM still halves the rate, even with a single subscriber.

DDS configuration

On the robot (CPU) — cyclonedds.xml

<?xml version="1.0" encoding="UTF-8" ?>
<CycloneDDS xmlns="https://cdds.io/config"
            xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
            xsi:schemaLocation="https://cdds.io/config https://raw.githubusercontent.com/eclipse-cyclonedds/cyclonedds/master/etc/cyclonedds.xsd">
    <Domain id="0">
        <General>
            <Interfaces>
                <NetworkInterface name="enp2s0" multicast="default" />
            </Interfaces>
            <AllowMulticast>spdp</AllowMulticast>
        </General>
        <Internal>
            <SocketReceiveBufferSize min="10MB"/>
        </Internal>
        <Discovery>
            <Peers>
                <Peer address="192.168.33.198"/>
                <Peer address="192.168.33.197"/>
            </Peers>
        </Discovery>
    </Domain>
</CycloneDDS>

On the VM — cyclonedds.xml

<?xml version="1.0" encoding="UTF-8" ?>
<CycloneDDS xmlns="https://cdds.io/config"
            xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
            xsi:schemaLocation="https://cdds.io/config https://raw.githubusercontent.com/eclipse-cyclonedds/cyclonedds/master/etc/cyclonedds.xsd">
    <Domain id="0">
        <General>
            <Interfaces>
                <NetworkInterface name="enp0s3" multicast="default"/>
            </Interfaces>
            <AllowMulticast>spdp</AllowMulticast>
        </General>
        <Internal>
            <SocketReceiveBufferSize min="10MB"/>
        </Internal>
        <Discovery>
            <Peers>
                <Peer address="192.168.33.197"/>
                <Peer address="192.168.33.198"/>
            </Peers>
        </Discovery>
    </Domain>
</CycloneDDS>

Notes

• Discovery works fine — the VM sees the topics and connects.
• 3.15MB pointcloud published at 10Hz

Questions

• Is there something wrong or missing in my DDS configuration that could explain the degraded performance over Ethernet when subscribing from the VM?
• Why does <AllowMulticast>spdp</AllowMulticast> solve the problem on a single machine, but not when the subscriber is on a separate machine?
• Any advice on further tuning or on how to debug this? I am not a DDS expert, honestly.

Thanks for your support!