Industrial Switch Buffer Sizing: The Loss That Certification Cannot See
Quick answer / TL;DR: Short answer: a buffer absorbs bursts, not sustained oversubscription. If the average ingress rate exceeds the egress rate, the queue fills regardless of buffer size. Size the uplink above the aggregate, and treat buffer depth as the mechanism that covers momentary collisions rather than as a substitute for bandwidth.
Three data families: USB carries power too, Ethernet spans buildings, fiber spans kilometers.A switch that drops frames while every cable passes certification is describing a buffering problem, and the diagnosis is frequently misdirected at the cable because that is what the technician can test.
Buffering exists to absorb the difference between what arrives and what leaves over a short interval. When the average rate exceeds the egress capacity, no buffer size helps, because the queue grows without limit.
This article separates burst absorption from sustained oversubscription, gives the arithmetic for both, and shows the counters that identify which one you have.
A store-and-forward switch receives a frame completely before forwarding it, so frames from multiple ingress ports can arrive at once and contend for one egress port.
The buffer holds the losers of that contention. If the contention is momentary, the buffer drains and nothing is lost. If the contention is continuous, the buffer fills and stays full.
Burst Versus Sustained
| Condition | Ingress vs egress | Buffer outcome | Symptom |
|---|
| Momentary burst | Briefly above | Drains after the burst | No loss |
| Repeated bursts | Above for intervals | Drains between intervals if deep enough | Loss under load |
| Sustained oversubscription | Continuously above | Fills and stays full | Loss proportional to excess |
| Incast | Many to one at once | Fills quickly | Loss with short recovery |
Incast is the case that surprises people. Several ports send to one at the same moment, the egress port is momentarily the bottleneck despite adequate average capacity, and loss appears without any sustained oversubscription.
That pattern is common in machine vision, where several cameras trigger simultaneously on a hardware signal.
The Counters That Identify It
- Read egress queue drops, which count frames discarded because the queue was full
- Read ingress discards, which indicate the port itself could not accept the frame
- Compare against CRC errors, which indicate a physical problem rather than a buffering one
- Watch queue depth over time, because a queue that sits near full predicts the next loss
- Record all three at commissioning as a baseline
Real-World Applications
A vision station triggers four cameras on one strobe. The upload from each camera lands in the same interval, and the uplink drops frames despite an average rate well below capacity. Staggering the triggers resolves it without changing hardware.
A line with eight cameras and a gigabit uplink loses frames whenever all cameras run at full rate. The aggregate is 8 Gbps against a 1 Gbps uplink, which is sustained oversubscription and cannot be buffered.
A packaging machine shows periodic loss that correlates with a specific product change. The change increases frame size, which changes the burst profile rather than the average rate.
A plant network shows ingress discards on one port during shift changes, when a barcode scanner bursts. The port is at its rate limit momentarily, and the fix is the scanner's trigger rate rather than the switch.
Standards and Verification
Ethernet switching behaviour is defined in IEEE 802.1 and 802.3, with priority and flow control mechanisms in IEEE 802.1Q and 802.3x. GigE Vision transport is maintained by the Automated Imaging Association, which documents the resend behaviour that masks buffer loss.
Verification should include a sustained test at production rate with the queue and discard counters monitored, plus a burst test that reproduces the worst simultaneous trigger condition.
A test at idle proves nothing about buffering. The condition that fills a buffer is the condition that only occurs when the machine is doing its job.
Designing Around the Burst
Where simultaneous triggering is the cause, the cheapest fix is usually temporal rather than electrical.
Staggering camera triggers by a few milliseconds converts a burst that saturates an uplink into several smaller bursts that do not.
That change costs nothing, requires no hardware, and is frequently rejected because it is not a purchasing decision.
Why the Cable Gets Blamed
The cable is the one element a technician can test, so a passing test feels like an unfinished investigation.
The counters separate the two immediately. A clean link with rising egress drops is not a cable problem, and no amount of recabling will change it.
Read the counters first. They cost less than a cable and answer the question directly.
Where staggering is not possible because the application needs simultaneity, the answer is more egress capacity or a dedicated segment, not a deeper queue.
Simultaneity is a requirement to be met with bandwidth, not with storage.
Related reading: GigE Vision multi-camera bandwidth — blogs/news/gige-vision-multi-camera
Related reading: GigE Vision PoE budget — blogs/news/gige-vision-poe-budget
Related reading: Industrial Ethernet certification test — blogs/news/industrial-ethernet-certification-test
A deeper buffer moves the loss later; it does not remove it. When ingress exceeds egress on average, only more egress capacity fixes the problem.
Frequently Asked Questions
Why do frames drop when all cables pass certification?
Because certification measures the physical channel. Buffering and oversubscription are properties of the switch and the traffic pattern, not of the cable.
Does a bigger buffer fix packet loss?
Only for bursts. If the average ingress rate exceeds the egress rate, the buffer fills regardless of size.
What is incast?
Several ingress ports sending to one egress port at the same moment, which momentarily exceeds that port's capacity even when average utilisation is low.
Which counter identifies a buffer problem?
Egress queue drops, together with a rising queue depth. CRC errors indicate a physical problem instead.
Sourcing cables for a project, a store, or a private-label line? Contact our B2B sales team for specifications, certification evidence, and volume terms – FOB and CIF options available. Request a quote or sample.
