Mesa Ethernet / QtPlasmaC retrofit experience — lessons from a marginal mini PC

I wanted to share my experience from the last several days in case it helps someone else planning a Mesa Ethernet / QtPlasmaC plasma retrofit.

This is not a complaint about LinuxCNC, far from it. After working through the issues, I am still happy to be moving forward with LinuxCNC and QtPlasmaC because I think the plasma feature set is excellent. But I do think my experience shows that the “PC side” deserves more attention than I initially gave it.

My setup:

- Home-built plasma table conversion from MASSO to LinuxCNC / QtPlasmaC
- Mesa 7I76EU
- THCAD2 planned
- Ethernet Mesa connection
- QtPlasmaC
- Separate camera-assisted project running alongside LinuxCNC during some tests

One of my reasons for moving away from a closed standalone controller was that I wanted to experiment with a camera-assisted tracing/scanning workflow, somewhat similar in concept to SheetCam’s Scanything. I have been working on a small project I call FabScan, which uses a USB camera mounted to the machine to help trace parts/templates and eventually export geometry. That type of experimentation is much easier on an open PC/LinuxCNC system than it would be on a closed standalone controller such as MASSO.

I originally planned to use a small MeLe mini PC that I already had. It booted fine, LinuxCNC installed, QtPlasmaC launched, Mesa connected, and the machine homed and jogged. At first glance it seemed usable.

The problem was intermittent realtime/Mesa communication errors. I saw errors such as:

- Unexpected realtime delay
- hm2 error finishing read
- Watchdog has bit
- Smart Serial communication error / timeout
- Smart Serial Port 0 stopped

The PC had a Realtek RTL8111/8168 Ethernet controller using the r8169 driver. I tried the usual tuning steps:

- Dedicated Mesa Ethernet interface
- Static IP, no gateway, no DNS on the Mesa NIC
- Disabled Ethernet offloads
- Disabled EEE
- Disabled coalescing
- Disabled Wi-Fi during testing
- Disabled power-management options in Linux
- Increased servo period from 1 ms up to 3 ms and later 4 ms
- Switched from r8169 to r8168-dkms

The r8168 driver helped a lot. The system would run longer, but it still eventually produced realtime delay / watchdog / Smart Serial failures. In my case, the MeLe was simply not boring enough to trust as a control PC.

I also found a separate issue that was not the PC’s fault. My generated config had PID loops driving Mesa velocity-mode stepgens, with DEADBAND = 0.0 on the joints. After homing or jogging, the axes would twitch slightly because the PID loop was chasing sub-step position error. Adding deadband fixed that:

X/Y/Y2:
DEADBAND = 0.0008

Z:
DEADBAND = 0.00025

That stopped the axis twitching.

Eventually I replaced the MeLe with a used Dell OptiPlex 7060 Micro with Intel Ethernet. With the Dell, I tested QtPlasmaC, FabScan/camera load, and NoMachine remote desktop while watching Smart Serial. The Dell stayed clean at a 2 ms servo period. Smart Serial stayed at:

fault-count = 0
port_state = 3
run = TRUE

For over an hour, the takeaway for me:

LinuxCNC can run on modest hardware, but for Mesa Ethernet I would not say “any old PC” is good enough without testing. I would now strongly prefer:

- Used business-class Dell/Lenovo/HP mini PC
- Intel wired Ethernet
- Good BIOS power-management controls
- Dedicated Mesa NIC
- Long latency tests, not just short ones
- Real workload testing: QtPlasmaC, camera, remote desktop, Wi-Fi if used, etc.

I am not posting this to criticize LinuxCNC, again, far from it. QtPlasmaC is exactly why I am doing this conversion. I just think new users coming from MASSO, Mach, GRBL, etc. would benefit from clearer up-front guidance that the control PC needs to be “boring,” not just powerful enough.

In my experience, horsepower was not the issue so much as the details. The NIC, driver, BIOS/power-management behavior, and long-duration realtime stability mattered a lot. I wish I had understood those details better up front, so I am posting this in case it helps the next person avoid a few days of chasing intermittent gremlins.

I am far from a LinuxCNC expert, but I like to think I have decent troubleshooting skills & from what I saw the MeLe Realtek NIC was the biggest issue.. There may have been some other power-management issues as well, but it really looked like the NIC would drop out just long enough to cause a problem...

Realteks became a problem with Bullseye (5.10 kernel) but we never released a version on it so the first issues were found with Bookworm (6.x kernel) when the kernel NIC driver architecture changed.

I think the minimum spec is i5 (4 core ) with Intel NIC.

People became a bit blase with PC specs until the recent kernels as above, but when I started with Mesa you had to compile the kernel to get the required  PREEMPT_RT we took a lot more care with PC selection...