Troubleshooting
Why HMI point updates run slow on an Ethernet PLC link
Short answer
When a screen value updates every 10 s while twin machines update in 0.4 s, run the same read load test on both paths. A slow machine reading 2,500 words/s against 56,000 words/s, with clean link counters, points at the controller: here the installed CPU was a 374 where the project specified a 374+. Verify the physical CPU identity before tuning any scan setting.
On this page
- Why does a timer value update every 10 seconds on one machine and 0.4 seconds on its twins?
- Which stage in the HMI-to-PLC update path do I check first?
- Is the Ethernet link itself at fault?
- Does the installed CPU match what the project assumes?
- How do I measure the read throughput the link actually delivers?
- Can scan and tick settings close a 25x update gap?
- How do I swap the CPU and prove the fix held?
- Questions asked
A timer value that refreshes every 10 s on one machine while its twin machines refresh the same value in 0.4 s is a data-delivery fault, not a screen fault, and a few measurements localize it. This project — about 6,000 points over Ethernet to a GE 90/30 controller — was starved by the controller itself: the installed CPU was a 374 where the project specified a 374+. Fitting the specified CPU restored full-speed updates. The sequence below traces the path from screen back to controller so you can prove which stage limits your machine before you change any setting.
Why does a timer value update every 10 seconds on one machine and 0.4 seconds on its twins? #
A 25x gap between twin machines running the same project size means the slow one is being starved of data somewhere between controller and screen, and the read load test later in this sequence measures exactly where the starvation sits.
Both machines ran the same project over the same Ethernet path with the same scan settings, so configuration alone could not explain the difference. The lag was also uniform: every value on the slow screen was stale by roughly the same interval, not just the timer. Uniform staleness points at delivery rate. Freezing followed by a jump would point instead at loss and retry, which is a different branch with different checks.
Treat the working twins as the baseline for every measurement that follows. Same project size, same controller family, same network — any number that differs by an order of magnitude between such machines marks the faulty stage.
Which stage in the HMI-to-PLC update path do I check first? #
Check the chain in a fixed order — HMI computer, Ethernet network, controller Ethernet port, controller data service, returned point data, screen rendering — because a slow screen on its own never tells you which stage is limiting delivery.
Work the stages in this order and let each one exonerate itself before moving deeper:
- HMI computer and driver: the runtime healthy, no other process loading the station.
- Ethernet path: link status and negotiated speed at both ends, error counters on any managed switch ports.
- Controller Ethernet endpoint: the same counters on the PLC side of the cable.
- Controller data service: how fast the CPU answers read requests end to end — the read load test measures this stage.
- Returned point data: quality flags and timestamps on the values the driver receives.
- Screen rendering: whether the runtime redraws unrelated objects on time.
On this machine stages 1 through 3 came back clean and identical to the twins, which moved the whole investigation to stages 4 and beyond — and stage 4 is where the fault lived.
Is the Ethernet link itself at fault? #
Read link status and error counters at both ends of the cable, and on any managed switch ports in between, before touching a single scan setting: clean, stable links move the suspicion off the network and onto the controller.
Compare the slow machine's physical path against a working twin rather than against a spec sheet. Read negotiated speed and link status at the HMI interface and at the controller port, and pull error and discard counters from any managed switch in the path. Inspect terminations and patch leads, because a marginal link can pass traffic while silently retrying.
- Rising error counters, link flaps or a speed and duplex mismatch against the twins: fix the path first, then repeat the measurement before judging anything downstream.
- Stable links, clean counters, identical negotiation on both machines: the physical layer is exonerated; carry the suspicion to the identity check below.
A different symptom belongs to a different branch. When values drop out or freeze rather than lag uniformly, the pattern is loss and retry, and Diagnosing intermittent HMI to PLC Ethernet comms hangs works that branch in detail. If your panel hangs on a serial link instead of Ethernet, Why an HMI freezes or updates slowly over serial applies the same capacity logic to that medium.
Does the installed CPU match what the project assumes? #
Read the controller identity from the module nameplate and from the identity the connected programming software reports, then compare both with the project documentation and a working twin: here the slow machine carried a 374 where its project specified a 374+.
Similar model names are not a comparison criterion; the exact module is. On this machine the nameplate and the software agreed with each other and disagreed with the project: the installed module was one level below the specified one, on hardware otherwise identical to its twins. Record the installed hardware and software identity before you compare any further settings, so the comparison runs between equal facts.
Keep a hardware difference as an explicit branch rather than an instant conclusion. The mismatch becomes the leading suspect, but the measured read test in the next section is what justifies spending the module: a name difference with equal measured throughput would point back to configuration instead.
How do I measure the read throughput the link actually delivers? #
Open the diagnostics utility under the equipment folder in the HMI software, connect to the controller and run the read load test with identical selections on both machines: the slow machine returned 2,500 words/s against 56,000 words/s on a working twin.
Use comparable machine states on both runs — same screens open, same active clients — so the difference reflects the path and controller, not the workload. The result is the driver's raw read throughput from the controller:
| Measurement | Slow machine | Working twin |
|---|---|---|
| Project size | about 6,000 points | about 6,000 points |
| Read load test result | 2,500 words/s | 56,000 words/s |
The arithmetic ties the two symptoms together:
- Throughput gap: 56,000 ÷ 2,500 = 22.4 — the working twin returns 22.4 times the data per second.
- Display gap: 10 s ÷ 0.4 s = 25 — the slow screen lags 25 times more.
Both ratios land in the same order of magnitude, and the transport gap accounts for the display gap with a little scheduling on top. The limiting stage is the read path ending in the controller's data service, not the screen rendering. One caution: do not divide 6,000 points by 2,500 words/s to compute a refresh time — points occupy different word counts and requests are grouped, so that division has no meaning.
Decisions from this test:
- Throughput an order of magnitude low, link counters clean: controller branch — verify the installed identity and replace to the specified model, then retest.
- Throughput comparable to the twins but screens still slow: point acquisition and display branch — compare scan grouping and screen objects; the link and controller are innocent.
- Both machines read slow: the baseline itself is wrong — compare device configuration and active client load against a known-good project before judging either machine.
Can scan and tick settings close a 25x update gap? #
No — identical settings cannot produce a 25x gap between twins, and here both machines ran the same values, an Ethernet port setting of 25 ticks and a point scan time of every 1 tick, so the settings were exonerated before the CPU was.
A tick is a project-defined interval, not a fixed number of milliseconds: the installed software's help defines the duration and the exact function of each tick field, so never convert a tick count to seconds by rule of thumb. What the settings express is demand — a point scan time of every 1 tick asks for values at that pace, and the Ethernet port value governs how requests are paced. A requested interval does not guarantee fresh data at that interval. When the controller cannot service the demand, shortening the interval adds load without freshening a single value; the screen keeps showing whatever the data service managed to return.
If you suspect settings anyway, compare projects item by item: device configuration, point acquisition rates, address grouping, active clients and controller workload. Equal point counts generate unequal traffic — contiguous addresses can be collected in grouped requests, while scattered ones cost a transaction each. Correct any material difference you find against the twins, repeat the read load test, and let the measurement close the branch. On this machine the settings matched the twins exactly, so the branch closed and the confirmed identity plus the low throughput carried the controller decision.
How do I swap the CPU and prove the fix held? #
Replace the module with the specified model at a controlled stop, change only the CPU in that test and re-run the same read load test afterwards — the 374-to-374+ swap here restored point updates to the speed the twins already showed.
One safety point before anything else: a CPU swap reboots the controller, so the machine must be in its approved maintenance state with motion stopped and outputs in a safe state before the module comes out.
- Back up the controller application, the hardware configuration and the HMI project, and record the communication settings plus the baseline words-per-second result.
- Check the replacement module's documentation for application, rack, I/O and software compatibility, and complete any required conversion work before the shutdown, not during it.
- Place the machine in its approved maintenance state and change the module per the manufacturer's procedure.
- Restore the application and configuration, then confirm the reported identity, operating state, I/O status and Ethernet destination on the restored controller.
- Reconnect the HMI, check point quality, and re-run the read load test with the same selections used for the baseline.
Keep the test single-variable: do not adjust scan rates, groupings or network settings in the same shutdown as the module change, or the improvement cannot be attributed and rollback becomes guesswork.
The final decision for this class of fault: clean link counters plus a read rate an order of magnitude below a twin plus an installed CPU below the specified model means replace the CPU to the specified model and retest — which is exactly what restored full-speed updates on this machine. That outcome is a demonstrated correction for this machine, not a throughput specification for every module carrying either name. If the retest with the correct module still reads low, the branch moves to network configuration and active client load with the identity question closed.
Questions asked
Why do two 90/30 CPU names that look alike perform so differently?
The suffix is the specification, not decoration. Within this CPU family the plus variant services communication requests substantially faster than the base module: here the difference measured 2,500 words/s against 56,000 words/s on identical projects. Read the identity from the module nameplate and from the connected programming software, and record it before comparing two machines.
Will a project with more points always update more slowly?
Point count sets demand, not the ceiling on delivery. Two machines here each ran about 6,000 points with the same scan settings, and the faster controller still returned 56,000 words/s. What matters is the controller's communication service capacity, how many words each point occupies, and how scattered the addresses are. Judge a project by a measured read test, never by the point total alone.
My HMI logs PLC communication timeouts instead of slow values. Is that the same fault?
No. A timeout means requests or replies are being lost or never answered, so look first at addressing, cabling, switch load and any recent network changes. The fault described here delivered every value, just late: clean counters, no timeouts, and throughput an order of magnitude low. Slow-but-complete data points at controller service capacity; missing data points at the path.
Why did my read load test return a different number on a second run?
The test measures what the driver can push at that moment, and active clients change it. Extra open screens, a second operator station, a historian or an engineering laptop polling the same controller all consume service capacity. Close or note the active clients, keep identical test selections and compare only runs taken under the same machine state.