How-to
Segregating machine drive traffic from the plant network
Short answer
Put every machine node on managed switches with IGMP snooping and exactly one active querier per machine VLAN, and make the plant path cross a Layer 3 routing hop — a Layer 3 switch or a NAT firewall — because a VLAN tags traffic but never routes between 192.168.1.x and 10.10.1.x. Hold multicast-only drives such as the PowerFlex 40 at 100 ms RPI: its serial DSI back-end caps updates near 35 ms.
On this page
- Why does drive multicast flood the plant network?
- How much multicast traffic does a machine cell actually generate?
- Does IGMP snooping stop the flood by itself?
- Can a VLAN alone connect 192.168.1.x to the 10.10.1.x plant?
- How fast can a PowerFlex 40 update over Ethernet?
- Can part of the machine sit on an unmanaged switch?
- How do you commission the segregation step by step?
- What happens to the machine when the panel power drops?
- Questions asked
The flood that makes a machine network look unreliable is multicast drive traffic, and the fix is topology plus two switch settings, not more bandwidth. Put every machine node — drives, controller, HMI, commissioning ports — on managed switches with IGMP snooping enabled and exactly one active querier on the machine VLAN. Put the plant on the far side of a real routed hop, a Layer 3 switch or a NAT firewall, because a VLAN tags frames but never routes between 192.168.1.x and 10.10.1.x. Hold multicast-only drives such as the PowerFlex 40 at a 100 ms RPI: the drive's serial DSI back-end cannot serve Ethernet updates much below 35 ms no matter how clean the network is. The sections below give the arithmetic, the three boundary options and the commissioning steps to hand a technician.
Why does drive multicast flood the plant network? #
A controller's I/O connection publishes drive status as IP multicast unless both ends are configured for unicast, and drives without a unicast option keep it multicast forever, so a switch without snooping copies every frame to every port.
An implicit I/O connection works as a publish-and-subscribe pair: the controller sends its outputs, and each drive answers by publishing status to a multicast group address. Where both ends support unicast you can convert the exchange to point-to-point frames, but the PowerFlex 40 has no unicast option, so its status stream is multicast by design.
An unmanaged switch cannot know who wants that stream, so it treats multicast like broadcast and replicates every frame out every port — including the port the plant electrician patched in last week. Every HMI, laptop and plant-side host on the segment then receives all of the drive streams and discards them in software. Devices with small processors stall first, and to an operator it looks like random communication loss that begins the moment the plant cable goes in.
How much multicast traffic does a machine cell actually generate? #
Budget roughly 0.2 Mbit/s of multicast for eight drives at a 100 ms RPI — the load is trivial, and that is exactly why the damage comes from flooded copies rather than bandwidth.
Take the machine at hand: eight small drives, three larger ones, one controller, one HMI, two commissioning ports and one plant uplink — sixteen ports.
- One status frame per RPI per drive: at 100 ms that is 10 frames per second.
- Assume 150 bytes per frame on the wire, a fair size for a small implicit I/O frame with its IP and UDP headers; this assumption moves the total by a factor of two at most, never an order of magnitude.
- Command plus status per drive: 2 × 150 B × 10/s = 3,000 B/s = 24 kbit/s.
- Eight multicast drives: 8 × 24 kbit/s = 192 kbit/s, about 0.2 Mbit/s on a 100 Mbit/s link — roughly 0.2 percent of the wire.
Now the unmanaged case. Eight drives produce 80 status frames per second, and a 16-port switch without snooping copies each one to the other 15 ports: 80 × 15 = 1,200 copies per second, about 1.44 Mbit/s of waste. The bandwidth still fits easily — the harm is that the HMI, the programming laptops and the plant uplink each process 80 unwanted frames every second, and small processors are the ones that bog down. Snooped correctly, the same 80 frames per second go to the controller port alone.
This arithmetic settles the prioritization question too: there is no bandwidth competition to arbitrate on a machine link carrying under 1 Mbit/s, so QoS classes and priority VLANs buy nothing here. Spend the effort on snooping and the routed boundary.
Does IGMP snooping stop the flood by itself? #
No — snooping prunes only the groups it sees joined through IGMP, memberships age out on a timer, and without an active querier refreshing them the switch reverts to flooding; run exactly one querier per machine VLAN.
Snooping is the switch reading IGMP join and leave messages and forwarding each multicast group only to the ports that asked for it. The querier is what keeps it true: a membership expires if nobody refreshes it, so one device must send the periodic IGMP general queries that make every listener re-announce itself. Enable the querier on the machine VLAN, on exactly one device — two queriers in one broadcast domain trigger an election fight, and the symptom is intermittent flooding that reads as random drive communication loss. The settings that matter and the counts to check are laid out in IGMP querier requirements for EtherNet/IP control networks.
Plan for the cold start as well. Snooping does not prune instantly: on some switches the pruning process takes minutes after power-up, and until it completes, multicast floods every port. Keep hosts that cannot tolerate flood noise off the machine VLAN instead of tuning around the start-up window.
The classic mistake — snooping on, querier forgotten — produces timeouts rather than obvious floods: controller connections drop and HMIs freeze while the link looks healthy. That is the same pattern documented in a ControlLogix connection timeout fixed by adding the missing IGMP querier, so check the querier before you replace a single drive.
Can a VLAN alone connect 192.168.1.x to the 10.10.1.x plant? #
No — a VLAN tags frames at Layer 2 and never routes, so no switch configuration alone connects 192.168.1.x to 10.10.1.x; you need one routed hop, and a Layer 3 switch or a NAT firewall is the standing choice for this boundary.
A VLAN separates broadcast domains, nothing more. The device that routes between the subnets is also your multicast wall: a router does not forward IP multicast unless multicast routing is configured, and you will not run multicast routing for control I/O. Build the boundary one of three ways.
| Boundary choice | What it gives you | What to watch |
|---|---|---|
| Layer 3 switch at the machine edge | Routes both subnets in one box and keeps full port counters in the machine panel | Boots in minutes; feed it from power that survives a panel isolation |
| NAT or port-forwarding firewall | Machine addressing stays identical on every build; only the listed ports cross to the plant | The forwarding list is also the security boundary; keep it minimal |
| Renumber the machine into the plant scheme | No new hardware at the boundary | Puts drive multicast back into the plant broadcast domain; safe only where the plant already runs snooping with one querier everywhere |
For a machine whose small drives are multicast-only, the decision is the Layer 3 switch or the NAT firewall; do not renumber. A DIN-rail industrial firewall belongs in the machine panel, and keeping machine addressing identical across every shipped unit is worth as much as the routing itself. Treat the crossing deliberately and stack defenses there, the way layered OT security controls protect a network path. If the plant side is a SCADA host reached from outside the plant, shape that access like remote SCADA access through firewalls and proxies.
How fast can a PowerFlex 40 update over Ethernet? #
About 35 ms is the floor: the PowerFlex 40 bridges Ethernet to its power stage through a serial DSI port, so set its RPI at 100 ms and never use this drive family in a fast Ethernet control loop.
The limit is the drive, not the network. The adapter serves the drive over that serial link, which cannot feed useful updates much faster than the figure above; a 100 ms RPI carries roughly three times that as margin, and anything faster buys nothing. Keep the family on non-demanding duty — in applications like this it was never time-critical anyway.
The larger drives are a different story. A PowerFlex 755 serves its ordinary Command / Reference / Status / Feedback connection through the DPI port and responds to changes on the order of 10 ms, comparable to the older drive families and enough for semi-time-critical duty such as coordinated speed and indexing. Where both ends support it, configure those connections as unicast and their streams leave the multicast set entirely. Safe torque-off capability is an independent reason to spec the 750 family. If a unit has to run as a true motion axis, that is CIP Motion, which brings its own controller-family and switch-model requirements — confirm the platform with the drive vendor before you quote it; ordinary drive connections need none of that.
Can part of the machine sit on an unmanaged switch? #
None in the I/O path: an unmanaged switch re-floods multicast to all of its own ports and leaves you with no counters, so buy one managed switch with at least sixteen ports instead.
An unmanaged switch downstream of the managed one is invisible to IGMP, so it does not disturb the single-querier rule — but it undoes the pruning for every port behind it. Each of the eight small drives then receives the seven streams published by its neighbors on top of everything else on the segment, and end devices with small processors are the ones that overload. You also lose the port counters and port disable that make the machine diagnosable; a replacement drive that never appears in the controller's device list is far harder to trace without them.
Count ports instead of splitting the switch: eight small drives plus three larger drives, one controller, one HMI, two commissioning ports and one plant uplink is sixteen. Buy one managed switch of sixteen ports or more, or cascade two managed switches over a tagged trunk with snooping enabled on both and the querier count unchanged.
How do you commission the segregation step by step? #
Eight steps take a technician who knows PLCs and DeviceNet through the whole job: draw the boundary, assign the VLAN, enable snooping, enable one querier, set the RPI, then verify counters, ping across the hop, and time the reboot.
- Draw the addressing plan: machine nodes on 192.168.1.0/24 (mask 255.255.255.0), plant on 10.10.1.0/24, and one routed device chosen for the boundary — a Layer 3 switch or a NAT firewall.
- Assign every machine node to the machine VLAN and tag that VLAN on the uplink to the boundary device.
- Enable IGMP snooping on the machine VLAN on every switch in the machine chain.
- Enable the IGMP querier on exactly one device in that VLAN and record which one.
- Set the RPI in each connection's properties: 100 ms on every multicast-only drive, unicast on the larger drives where supported.
- Verify pruning: at 100 ms each drive port should show about 10 multicast packets per second, and the HMI and plant uplink ports should show none of the drive groups.
- Prove the boundary: ping the controller from a plant-side PC through the routed device, then confirm that no drive multicast counter moves on the plant uplink.
- Time a cold reboot of the boundary device with a stopwatch, record the figure, and decide its power feed from it.
If step 6 or 7 shows drive communication faults instead of clean counters, suspect the snooping configuration before the hardware — that is the pattern behind a PowerFlex 755 F5037 fault traced to IGMP snooping setup on Stratix 8000 switches. None of the eight steps needs deep Ethernet experience: the switch settings live in the switch's management pages, the RPI lives in the controller's connection properties, and the checks are counter reads.
What happens to the machine when the panel power drops? #
Plan for minutes, not seconds: budget a 3 minute boot for a chassis-class Layer 3 boundary device, and remember that any interlock carried across the plant link is dead for the whole window whenever that device loses power.
The access switches return in far less time; the routed device is the one that takes minutes, and some chassis-class units run past five, so time your actual model and write the number on the panel drawing. Every process interlock that crosses the plant link is unavailable for that entire boot window after a power loss.
That leaves two decisions, and they are not exclusive. Feed the machine switch and the boundary device from control power that does not drop when the machine panel isolates — a dedicated feed, a UPS or dual sources — and the boot window never appears in production. Or design the machine sequence so that a dead plant link is a known, safe operating state and the boot is simply an outage you planned for.
Personnel-safety functions stay hardwired and independent of every switch and router in this article: no Ethernet path described here may ever carry a safety function.
Questions asked
Does IGMP snooping contain broadcast traffic too?
No. Snooping prunes only multicast groups that devices join through IGMP. Broadcast frames such as ARP still reach every port in the VLAN, which is harmless inside a /24 machine subnet with a dozen nodes, and it is one more reason the plant path should cross a router: routers stop broadcasts the same way they stop multicast.
Can two controllers share the same machine VLAN?
Yes. Each controller joins the drive multicast groups on its own behalf, snooping forwards each stream to every member port, and the single querier serves both. Set each connection's RPI independently and keep the 100 ms floor on every multicast-only drive no matter which controller talks to it.
What IP addresses should the machine cell use?
Any private /24 the plant does not already use. The examples in this article put the machine on 192.168.1.0 with mask 255.255.255.0 and the plant on 10.10.1.0; avoid any range your remote-access VPN hands out, and write the chosen range into the machine documentation so a future build does not collide.
How does the plant side get drive data for logging?
Not by joining the drive multicast. A routed boundary blocks multicast by design, and you should not enable multicast routing for control traffic. Let a plant-side or edge host read the controller over unicast across the routed path and log from there; the drive streams never leave the machine VLAN.