IMD SolutionComputing LibraryShop main store

Troubleshooting

Why OneDrive sync does not belong on a machine controller

Operating systems on the plant floor Published 8 min read

Short answer

Run the OneDrive sync client on an office PC, not on the machine controller. A TwinCAT PC reserves cores for real time, and a sync client adds burst CPU, disk and network load on the Windows side that shows up as task exceeds and EtherCAT lost frames. Sync the SharePoint library to a separate PC and let the controller read and write the folder over a file share.

On this page
  1. What actually breaks when a sync client runs on the controller?
  2. Does the controller's Windows image support the sync client?
  3. Where does the sync traffic travel?
  4. Will the sync load reach the real-time tasks?
  5. How do you measure the impact before committing?
  6. Where should the sync client run instead?
  7. What must pass before you release the machine?
  8. Questions asked

Run the OneDrive sync client somewhere else. On a machine controller running TwinCAT, every add-on competes with the real-time side for CPU time, memory bandwidth, disk throughput and interrupts, and a sync client loads all four in bursts: the initial library download first, then every bulk edit someone saves to SharePoint. Field reports on this class of install describe task exceeds and EtherCAT lost frames appearing whenever a sync triggers. The fix is architectural, not a tuning job: sync the library to an office PC or server and let the controller read and write the folder over a plain file share. Work through the checks below in order; each one either closes the question or points to the branch that fits.

What actually breaks when a sync client runs on the controller? #

The sync client never touches a reserved real-time core directly, but it loads the Windows side in bursts of file hashing, TLS work, disk writes and file-system notifications, and the two sides still share memory bandwidth, cache, the storage controller and the interrupt subsystem.

TwinCAT runs its tasks on cores reserved from Windows, and that reservation separates scheduling only. Client activity on the Windows side therefore lands on hardware the real-time tasks depend on too. The initial download of a whole library is the worst case; incremental syncs are smaller but repeat all day. The symptoms seen in the field are:

Symptom Where to read it Mechanism
Task exceeds Exceed counter and minimum and maximum cycle time of each task in the TwinCAT online views Shared-resource contention or long interrupt activity while a sync burst runs
EtherCAT lost frames Lost-frame counters in the EtherCAT master and slave diagnostics A cycle overrun delays the frame send, or the master misses its receive window
Both, only during sync bursts Timestamps correlated with the client's own activity log The sync burst is the trigger; removing it from the controller is the fix

Real-time impact from added software cannot be read off a datasheet. Measure it, using the procedure further down; the checks in between establish whether the install is possible at all.

Does the controller's Windows image support the sync client? #

Install it and the client usually starts, but four checks decide the branch: the client's own requirements (a 1.6 GHz or faster processor and at least 1 GB of RAM), free disk space for the library, the tenant's sync policy, and the trimmed image itself.

  • Build. Record the exact Windows build with winver and compare it against Microsoft's OneDrive system requirements, which also name a minimum display resolution that a headless unit can trip over. A build the client no longer supports closes the local branch at once.
  • Image. Controller vendors ship trimmed Windows images, and field reports describe the client installing but refusing to run on stripped LTSC setups. Treat that as a possible outcome of your build, not a documented behavior of any specific one.
  • Disk space. Without on-demand files, the whole library mirrors to the local drive: a 12 GB library needs 12 GB free before the first sync. Set the target at library size plus 20 %, which is 14.4 GB for that example, because a full system drive also breaks Windows and TwinCAT logging.
  • Tenant policy. The Microsoft 365 admin can block sync clients on unmanaged or non-domain devices outright. Confirm the policy before spending a shift on benchmarks.

All four passing means the install is merely possible; the network path and the load decide the rest. Any failure here sends you to the off-box branch below.

Where does the sync traffic travel? #

Sync traffic is outbound HTTPS from the Windows-owned network port to Microsoft 365, so the path is controller, plant switch, firewall, internet — and a port bound to the TwinCAT real-time Ethernet driver cannot carry any of it.

Identify the adapters before anything else: when the real-time Ethernet driver claims a port, that adapter disappears from Windows and cannot be used, so the client must live on the port Windows keeps. Then walk the hops:

  • VLAN. Put sync on an office VLAN, not beside fieldbus and HMI traffic on the machine network. That separation is one case of the reasoning in why layered OT security controls protect a network path.
  • Firewall. Outbound HTTPS to the tenant's Microsoft 365 endpoints must be permitted, or the client signs in and then never moves a file.
  • Proxy and TLS inspection. Either appliance can break the sign-in handshake; a proxy or inspection box is a common blocker that looks like a dead client.

A clean path leaves only local load, which is the check that actually decides the question.

Will the sync load reach the real-time tasks? #

Read the core split first: TwinCAT reserves cores for real time, and on a core shared with Windows the real-time share can be capped anywhere between 10 and 90 percent, which decides how much desktop-grade CPU the sync client will ever see.

The real-time settings in the TwinCAT development environment show which cores Windows keeps and which are reserved, and each real-time kernel is scheduled independently, so a bad split stays invisible until a burst arrives. Reserving a core does not reserve the memory bus, the cache or the storage path behind it.

Worked example, on an illustrative unit: a two-core controller reserves core 1 exclusively and caps shared core 0 at a 70 % real-time share, so Windows keeps 30 % of one core. The sync client's documented minimum is a 1.6 GHz processor and 1 GB of RAM. On a unit with 1.6 GHz cores, the Windows side offers the client 1.6 GHz × 0.30 = 0.48 GHz — under a third of the client's own minimum — before Explorer, logging and antivirus take a slice. Memory: if the image and runtime hold 1.2 GB of a 2 GB unit, then 2 GB − 1.2 GB = 0.8 GB remain, below the 1 GB minimum before the first file lands. On that split the local install starts in the red and should be expected to fail the burst test.

An exclusive-core layout or a larger machine changes the budget, which is exactly why the next section measures instead of predicting.

How do you measure the impact before committing? #

Measure instead of predicting: run a baseline with the production cycle and no client, install on a test unit, record the same counters through an initial full-library sync and through incremental edits, and fail the local install if either phase moves them.

Keep the install off the production machine: do the steps on a test unit with the same image, build and project, and open a controller only inside a maintenance window with the machine in a safe state.

  1. Record a baseline on the real workload with no client installed: each task's exceed counter and minimum and maximum cycle time, plus the EtherCAT lost-frame counters, over at least one full production cycle.
  2. Install the sync client on the test unit, signed in to the tenant, with the same library configured.
  3. Start the initial sync of the full library — the worst case — and record the same counters throughout.
  4. Let it settle, then add and edit files until several incremental syncs have run, and record again.
  5. Compare against the baseline. Any rise in exceeds or lost frames in either phase fails the local install on that machine.
  6. Zero change across both phases means the install is viable for that library size and that machine load; confirm it holds after one logged shift.

The same baseline discipline applies when TwinCAT runs on hardware it was not built for; it is laid out step by step in running TwinCAT XAR on a third-party industrial PC.

Where should the sync client run instead? #

Put the client on an office PC or server and let the controller read and write the synced folder over a file share: the controller then sees file input and output only when a program actually opens a file, with no background sync load of its own.

The mechanics are ordinary: the office machine runs the sync client and keeps the SharePoint library in its synced folder, that folder is shared over SMB, and the controller maps the share and touches files only when the program asks. Nothing client-related — no process, no auto-update, no telemetry — runs on the control box. The share and the client stay on the office network, so the controller's only new traffic is the SMB exchange itself, and the acceptance test below covers exactly that.

If the customer requires on-controller sync and the measurement above fails, the remaining branch is a hardware change: the same project moves to a machine with more cores and a higher TwinCAT performance level, which means re-licensing and an end to standardization with the rest of the fleet. If that trade is accepted, evaluate an industrial mini PC with enough cores to leave Windows a full core on top of the real-time reservation, and repeat the full measurement on it before release.

Decision: off-box by default; a local client only with a measured pass; upsizing only when on-controller sync is a hard requirement and the larger unit passes the same test.

What must pass before you release the machine? #

Accept a branch only when the counters hold: exceed counter, minimum and maximum cycle time and EtherCAT lost frames must match the recorded baseline through a full sync burst — or a full file-share transfer on the off-box branch — and still match after one logged shift.

  • Local or upsized install. With the full production cycle running, trigger the initial sync of the real library and a burst of incremental edits; the exceed counter, the maximum cycle time and the lost-frame counters must sit at baseline values throughout.
  • Off-box share. Copy a large batch of files between the controller and the share while the machine produces; the same counters must sit at baseline.
  • Disk margins. After sync settles, confirm the system drive still holds the margin set earlier, and the volume behind the share has room for the library's growth.
  • One full shift. Leave the counters logging across a shift with sync active, and accept only if they still match the baseline at the end.

Every line holding means release with the branch you chose. Any line moving means the sync still burdens the controller more than the machine tolerates, and the answer falls back to the off-box branch, because it is the only one that removes the client from the box entirely.

Questions asked

Can OneDrive be installed for all users so a service account can use the synced folder?

On Windows 10 Enterprise LTSC the sync client installs per user: it exists only for the account that installed it, and other logons find nothing in the Start menu or search. A service or machine account therefore cannot consume a folder a technician synced, which is one more reason the off-box file share is the dependable route.

If the controller meets OneDrive's system requirements, is the local install safe?

No. A 1.6 GHz processor and 1 GB of RAM qualify the client to run at all, not to coexist with reserved real-time cores. The shared-core CPU budget, the free disk space for the library and the measured burst test decide whether the install is safe, and each is specific to your image and project.

What if the workflow truly needs the sync client on the controller?

Upsizing is the remaining branch: move the same TwinCAT project to a controller with more cores and a higher performance level, install the client there, and repeat the baseline and burst measurements on that unit before release. More cores shrink the contention but leave the shared memory, storage and interrupt paths in place, so measurement still decides.

The client installs and signs in but no files move — where do you look?

Reproduce the sync from an office PC on the same path. If it works there, walk the hops the controller shares: switch, firewall, proxy. The usual culprits are outbound HTTPS to the tenant's endpoints not yet permitted and a TLS-inspection appliance interfering with the handshake, and both live in IT's configuration, not on the machine controller.

Does Files On-Demand change the disk-space math?

Yes. With on-demand files enabled the client keeps placeholders on the controller and downloads content when a file is opened, so a 12 GB library does not need 12 GB free up front. The trade: every first open becomes a burst download, so on-demand swaps disk headroom for more frequent network and disk activity during operation.

Keep reading

  • How-to

    Running TwinCAT XAR on a third-party industrial PC

    Do not clone a factory IPC image onto other hardware. Install a licensed Windows 10 IoT Enterprise LTSC edition, install the runtime fresh, and qualify the real-time Ethernet port, CPU priority and jitter on the actual machine. A successful boot proves nothing, and network hardware the runtime vendor did not sell you is a self-supported component.

    Operating systems on the plant floor

  • Troubleshooting

    Fixing a historian device shortcut that will not appear

    A ping, an open web page and an online programming session prove IP reachability, not a working data path. When the historian drops a device shortcut, the RSLinx Enterprise tree usually holds the gateway interface instead of the controller behind it. Stop browsing and the related services, delete the interface entry, define the SLC 5/03 manually under ENI Devices at the same IP address, point a fresh shortcut at the controller icon, restart, and confirm with a live item read.

    Edge computing and IIoT

  • Checklist

    Proxmox for SCADA and PLC engineering virtual machines

    Yes — Proxmox runs Siemens SCADA and PLC engineering VMs; field teams use it for TIA Portal engineering over RDP and in-cabinet dedicated hosts. Commit per workload: pass the support-position, Layer 2 bridge, license and restore tests below before production, and size the host from measured demand.

    Remote access and virtualization