Not a LinuxCNC install. Same class of box as the earlier J1900 eglfs vs Xorg thread.Posting because the useful knob was not another kernel cmdline flag.
Intel Celeron J1900 has 2 MiB L2 as two 1 MiB slices:
CPUs 0-1 share one L2
CPUs 2-3 share the other L2
We already isolated the EtherCAT loop on CPU3 (SCHED_FIFO 80, 1 kHz).
That is not enough if the fullscreen Qt/eglfs HMI is still allowed on CPU2.
The UI then sits on the L2 sibling of the RT core and contends for the same 1 MiB.
Change (HMI only, not a whole-system cgroup):
before: Qt main / touch / QML / render / GPU worker allowed on CPUs 0-2
after: those threads allowed on CPUs 0-1 only
RT thread stayed on CPU3 both times
Same binary, same recipe, same 1 ms period, 20 automatic press cycles each side.
Jitter = wake-to-wake CLOCK_MONOTONIC minus 1 ms, integer microseconds.
0 us means |error| < 1 us. This is period error, not latency-test / cyclictest.
All numbers below are the full captures (approach + press + retract), not a PID subset.
Absolute jitter, nearest-rank, all samples
UI on 0-2 (20 runs) UI on 0-1 (20 runs)
n 119159 119254
P50 0 us 0 us
P95 6 us 1 us
P99 13 us 1 us
P99.9 21 us 2 us
signed min/max -42 / +46 us -13 / +14 us
|j| >= 10 us 2983 8
|j| >= 20 us 198 0
|j| >= 30 us 22 0
|j|
Per-run P50 / P95 / P99 / P99.9 (|jitter|, us)
# UI 0-2 UI 0-1
1 1 / 9 / 14 / 23 0 / 1 / 1 / 2
2 0 / 2 / 9 / 19 0 / 1 / 1 / 2
3 0 / 5 / 12 / 21 0 / 1 / 1 / 2
4 0 / 6 / 15 / 20 0 / 1 / 1 / 2
5 0 / 7 / 15 / 20 0 / 1 / 1 / 2
6 0 / 1 / 1 / 2 0 / 1 / 1 / 2
7 0 / 6 / 14 / 20 0 / 1 / 1 / 2
8 0 / 7 / 15 / 20 0 / 1 / 1 / 2
9 0 / 4 / 13 / 19 0 / 1 / 1 / 2
10 0 / 6 / 16 / 26 0 / 1 / 1 / 2
11 0 / 3 / 11 / 22 0 / 1 / 1 / 3
12 1 / 9 / 15 / 23 0 / 1 / 1 / 2
13 1 / 9 / 14 / 26 0 / 1 / 1 / 2
14 1 / 9 / 15 / 32 0 / 1 / 1 / 2
15 0 / 8 / 14 / 19 0 / 1 / 1 / 2
16 1 / 9 / 14 / 26 0 / 1 / 1 / 2
17 0 / 1 / 6 / 13 0 / 1 / 1 / 2
18 0 / 6 / 13 / 19 0 / 1 / 1 / 2
19 0 / 1 / 5 / 15 0 / 1 / 1 / 2
20 0 / 1 / 1 / 2 0 / 1 / 1 / 2
After the pin, every run has P99 = 1 us. 15/20 runs have max |j|
The 8 remaining >=10 us samples are two ticks each in runs 1, 8, 15, 16.
Point of the numbers
- P50 was already 0 us. The HMI was not ruining the median.
- The L2 sibling shows up in the tail: P95 6->1, P99 13->1, P99.9 21->2.
- Moving i915 IRQ off CPU2 was not sufficient by itself. User-space Qt
render can still land on CPU2 and thrash the RT core's L2.
- Do not set a cgroup AllowedCPUs=0-1 on the whole process: the FIFO
thread must still be able to bind to CPU3.
- Do not park every system thread on 0-1. This was only the HMI/render path.
Box (for context, not the variable under test)
- Debian 13 PREEMPT_RT 6.12, Qt eglfs_kms, no X
- IgH EtherCAT, 1 kHz, Lost frames = 0
- isolcpus / rcu_nocbs on CPU3, i915 IRQ on CPU0
Happy to hear if anyone running a GL HMI on Bay Trail saw the same
CPU2/CPU3 L2 pairing.
