IMDHMIControls LibraryShop main store

Selection guide

Sizing one Ignition gateway for 100,000 OPC UA tags

SCADA and data logging Published 8 min read

Short answer

A 100,000-tag OPC UA application fits one Ignition gateway at the published large-project benchmark of 8 cores, 16 GB RAM and SSD storage, which assumes at most 10% of tags changing into history and up to 75 concurrent clients. Qualify that server against your changed-value rate, historian load and client activity with a representative test, and distribute gateways on measured limits, not on tag count.

On this page
  1. The benchmark and what it assumes
  2. What actually loads the gateway
  3. Worked example: rate plan against the budget
  4. The selection
  5. Fix rates before buying cores
  6. Memory, hosting and redundancy
  7. One reported 150,000-tag installation
  8. The qualification test
  9. Questions asked

Yes, with one qualification to prove. Inductive Automation's server sizing guidance puts a 100,000-tag project on a single gateway at its large-project benchmark: 8 cores, 16 GB RAM and SSD storage, assuming no more than 10% of tags changing into tag history and up to 75 concurrent clients. Select that, then qualify it with a representative test before go-live. Splitting the architecture is a response to a measured load, a network boundary or an isolation requirement — not to a tag count with five digits in it. If the SCADA package itself is still an open choice, settle that first; selecting SCADA software for mixed PLCs and SQL logging covers that step.

The benchmark and what it assumes #

The sizing guidance works in tiers: small systems run on about 2 cores and 4 GB RAM, medium systems on 4 or more cores with 8–16 GB, and the large-project benchmark for 100,000 tags is 8 cores, the RAM figure above and SSD storage. These are starting configurations from the vendor's guide, and the guide is explicit that faster polling and additional features move the result.

The 8-core figure carries assumptions: at most 10% of tags changing into tag history, and no more than 75 concurrent clients. Both are load statements in disguise. The 10% figure is a changed-value and history-write budget, and 75 clients is a rendering and transaction budget. Exceed either and you are sizing beyond the benchmark, whatever the tag count says.

Two variants matter. A cloud-hosted gateway at 100,000 tags is recommended at 8 vCPUs and 32 GB RAM, so use the cloud figure rather than substituting the on-premises RAM number. The same architectural material scales to roughly 150,000 tags and about 100 concurrent clients as core counts step up, with 16 cores appearing at that tier.

What actually loads the gateway #

Tag count is an inventory number, not a load. Four separate quantities set the size: acquisition requests leaving the gateway, changed values arriving per second, history records written per second, and client transactions. The one that decides server size is changed values per second:

changed values/s = tag count × changing fraction ÷ interval in seconds

The guide's own illustration is 100,000 tags polled at 1 second with 10% changing: 100,000 × 0.10 ÷ 1 s = 10,000 changed values per second. That is the workload the 8-core benchmark is built around, and it is the number to carry through the rest of this selection.

Two things the formula does not say. A changed value becomes a history record only when the tag's storage behavior says so, and acquisition traffic depends on OPC UA subscription behavior, not on this multiplication: a subscription can sample at the source, aggregate requests, and deliver on a publishing interval of its own.

Worked example: rate plan against the budget #

All-fast plant. 100,000 tags, every point at 1 second, 10% changing: 100,000 × 0.10 ÷ 1 s = 10,000/s. That sits exactly on the benchmark budget. Decision: one gateway at the benchmark configuration — and a qualification test, because there is no headroom.

Mixed plant. 30,000 indications at 1 second with 20% changing, plus 70,000 status and totalizer tags at 5 seconds with 5% changing: 30,000 × 0.20 ÷ 1 s = 6,000/s, and 70,000 × 0.05 ÷ 5 s = 700/s, for 6,700/s total. That leaves 3,300/s of headroom against the 10,000/s budget. Decision: single gateway, comfortably.

Fast plant. 100,000 tags at 250 ms with 10% changing: 100,000 × 0.10 ÷ 0.25 s = 40,000/s, four times the benchmark budget. Decision: do not buy an 8-core single gateway for this. Move everything that is not an indication to a slower group first; if the fast set genuinely needs 250 ms, step up to the 16-core tier or split acquisition across two gateways by plant area.

The selection #

Arrangement Starting configuration Choose it when
Single gateway 8 cores, benchmark RAM, SSD; gateway memory 10.5 GB on Windows, 13.6 GB on Linux 100,000 tags, at most 10,000 changed values/s into history, up to 75 concurrent clients
Cloud-hosted gateway 8 vCPUs, 32 GB RAM Same tag count on cloud infrastructure, using the cloud recommendation
Redundant pair Two nodes, each at the full single-gateway size Availability requires failover; the backup carries no share of the load
Distributed gateways Two or more gateways, each sized on its own tag count and rates Measured resource ceilings, plant or network boundaries, or isolation requirements

Every row starts from the same place: rates fixed by process requirement, the historian on separate hardware, and the qualification test below. Distribution enters when a measurement or a boundary demands it — a gateway per plant area when the networks or the isolation requirements split, not because the tag list is long.

Fix rates before buying cores #

Ignition's tag groups exist so each point executes only as fast as its process requires. Sort the inventory into fast indications and slow status, totals and diagnostics before compensating for acquisition demand with hardware; the mixed-plant case above is that exercise.

Per group, choose the acquisition mode. Subscribed mode reduces traffic by letting the source push changes, but a source pushing excessive updates can load the network; polled mode keeps network load predictable by asking on schedule. Default to subscribed for OPC UA sources whose change traffic you have measured, and polled where the schedule must be predictable.

On an OPC UA connection, the diagnostics show each subscription's requested publishing interval, the revised interval the server granted, the tag count and the monitored nodes. A revised interval slower than requested means the server renegotiated your rate, and publishing cadence is separate from the source's underlying sampling. Diagnose stale values at the subscription and the source before concluding the gateway lacks CPU. The server side owns half of that contract; connecting an S7-1500 to SCADA over OPC UA from TIA Portal shows where those server-side settings live for one common controller.

Rate settings are an operational choice — no safety function may depend on how quickly a SCADA tag refreshes.

Symptom Likely cause Deciding check
Updates arrive slower than configured Server revised the publishing interval, or source-side sampling is slower Compare requested versus revised publishing interval in the connection diagnostics
Low gateway CPU but stale values Bottleneck at the controller, OPC server or network Check source load and update delivery end to end
Network load climbs during process activity Subscribed mode delivering every source update Compare traffic against the count of changing monitored values
History falls behind while live values stay current Historian or database throughput Compare incoming records with completed writes and queue depth

Memory, hosting and redundancy #

The guide sets the gateway's maximum memory at 66% of host memory on Windows and 85% on Linux. On the benchmark's host that works out to ≈ 10.5 GB on Windows and ≈ 13.6 GB on Linux; on the 32 GB cloud host it is 32 × 0.85 ≈ 27.2 GB. Host RAM and the gateway's allocation are separate settings — a VM specification alone says nothing about the second one, and the operating system keeps whatever is left.

Keep the historian database off the gateway server at this scale: the benchmark assumes the gateway is not also the database server, and the installation below ran its database separately. If the gateway must be a physical machine in the control cabinet rather than a VM, an industrial mini PC is the form factor for that host, sized against the same core, memory and SSD numbers before you commit.

Redundancy is an availability decision. In a standard pair the backup stands by and carries no share of the live workload, so each node is sized at the full single-gateway figure. A pair shares nothing but your exposure to a single failure.

One reported 150,000-tag installation #

One reported installation — reported, not measured here — runs about 150,000 tags from two large controllers plus roughly 60 smaller devices at a dozen tags each. Each gateway VM has 8 cores and the benchmark RAM size, the gateway's maximum memory is set to 8 GB, the historian database is a separate server, and CPU sits around 5% with occasional spikes to 10% under 4–6 concurrent clients.

Read that correctly. At 8 cores, a tag count 50% above the benchmark can idle near 5% CPU when the communications are optimized and the client count is modest. The low figure comes from that project's optimization, not from tag count being irrelevant: it confirms the 8-core benchmark has margin, and it is not a license to skip the test.

The qualification test #

  1. Load all 100,000 tags against the real control system, not a tag import alone.
  2. Apply the production rate plan and drive representative change rates, including one startup or batch transition — simultaneous-change events, not just steady state.
  3. Enable history on exactly the tags that will carry it, and confirm record rates match the rate plan.
  4. Connect the expected clients on their production screens with trends and alarms active.
  5. Through the worst transition, watch gateway CPU, heap usage and history queue depth.

Decision rule: sustained CPU well away from saturation and a history queue that drains to zero confirm the single gateway. CPU that climbs toward saturation through every transition, or a queue that grows, means fix the rate plan first — then step up cores or split acquisition.

Questions asked

How many OPC UA tags can one Ignition gateway actually hold?

The vendor's large-project benchmark puts 100,000 tags on 8 cores, benchmark RAM and SSD, and the same architecture discussion scales to roughly 150,000 tags and about 100 concurrent clients as core counts step up. One reported redundant installation ran about 150,000 tags on 8-core gateway VMs at that RAM size, idling around 5% CPU. The real ceiling is changed values per second and history writes, not the tag count.

Does a redundant gateway pair double my capacity?

No. In standard redundancy the backup is a standby that carries no share of the live load, so each node must be sized for the full workload. Choose redundancy for availability and size each node as if it were the only gateway. Capacity problems are solved with rate plans, larger nodes or distribution, never with the second node.

What changed-value rate should I design to?

Design to the benchmark's 10,000 changed values per second: tag count times changing fraction divided by interval in seconds. 100,000 tags at a 1 second interval with 10% changing is exactly 10,000. If your rate plan lands above that, first move status and totalizer tags to slower intervals; if a genuinely fast set still exceeds it, step up cores or split acquisition across gateways.

How much memory should the gateway get on the benchmark host?

The sizing guide sets the gateway's maximum memory at 66% of host memory on Windows and 85% on Linux: about 10.5 GB and 13.6 GB respectively at the benchmark's host RAM. That allocation is a separate setting from the host's RAM, so copying a VM specification without the gateway allocation misses half the configuration.

Should the historian database share the gateway server?

At this scale, no. The benchmark figures assume the gateway is not also the database server, and the reported 150,000-tag installation ran its historian on separate hardware. A separate server, or at minimum a separate volume, keeps history writes from competing with acquisition and client traffic for the same cores and disks.

Keep reading

  • Troubleshooting

    Fixing CODESYS OPC DA memory growth stopping SCADA reads

    The reads stop because WinCoDeSysOPC.exe, a 32-bit OPC DA server process, climbs to its 2 GB memory ceiling over weeks and then stops serving tags; killing it only resets the clock. The resolution that held was installing the current, licensed CODESYS OPC DA server — 3.5.22.40 in place of the CODESYSV35SP10Patch2 install — and proving it through the old failure window.

    SCADA and data logging

  • How-to

    How to learn PLC, SCADA and DAQ from scratch

    Learn the three as one project, not three courses: pick one route, wire one real input, program a three-state sequence, exchange one Modbus register, bind SCADA tags and store a timestamped value. Pass the five checkpoint tests on the way; simulation alone never exercises terminals, polarity or reconnection.

    SCADA and data logging

  • How-to

    Read Citect SCADA 2018 R2 tags from Ignition via OPC DA

    Citect SCADA 2018 R2 ships a built-in OPC DA server: enable it under the Topology activity in Citect Studio, connect Ignition's OPC DA module to it, and the Citect variable tags appear for browsing and import with no OFS, no CtAPI and no project rewrite. OPC UA exists only in later Citect releases. When Ignition and Citect run on different PCs, configure DCOM security for the account the Ignition service runs under.

    SCADA and data logging