How NFC badge scanning throughput changes a door under load

At 08:52 the queue outside the west entrance ran back to the taxi rank. Inside, eight lanes were open and the readers were working perfectly. The vendor had quoted a read time of 200 milliseconds and was right about that, which is why nobody in the show office believed the queue had anything to do with NFC badge scanning throughput.

They were wrong, and the reason is that the read is a small fraction of what a lane spends on each person. The tap is fast. The transaction around the tap is not, and the transaction is what sets capacity.

How fast is the tap, really?

Fast enough to ignore. ISO/IEC 14443 governs contactless proximity cards operating at 13.56 MHz, with part 1 on physical characteristics published in 2000 and parts 2 to 4 covering the radio interface, initialisation and transmission protocol following in 2001. Part 1 is now in a 2018 edition. The standard describes cards intended to work within roughly ten centimetres of the reader antenna, and the activation and anticollision sequence completes in tens of milliseconds.

Against a person walking at 1.2 metres per second, that is nothing. If the whole cost of a badge read were the radio exchange, a single lane could pass tens of thousands of people an hour and no exhibition would ever have queued.

So the read time is a distraction in every specification conversation I have sat in. The question worth asking a supplier is what the reader does when the read fails, how quickly it tells the person, and how long the barrier takes to open and close.

Service time is what you actually have to measure

Service time runs from the moment a badge is presented at the reader to the moment the lane is free for the next person. Measure it with a stopwatch on forty consecutive people at a real door, in the peak, and it decomposes roughly like this.

Approach and present, 1.3 seconds. The person arrives, finds the badge, positions it. Badges on lanyards worn under coats add time here, and so does anything the person is carrying.

Radio exchange and decision, 0.3 seconds. The read itself, plus the panel deciding whether the credential is entitled to that door.

Gate release and passage, 1.4 seconds. The barrier opens, the person walks through, the barrier resets. Fixed by the hardware and nearly impossible to improve without replacing it.

Recovery, 1.2 seconds as a mean across everybody. Most people pass first time and contribute zero. Roughly one in six taps fails, and those people spend five or six seconds re-presenting, looking for a steward and trying again, which averages out across the sample.

Total, 4.2 seconds. Three of those four components involve no radio at all.

Sampling matters more than sample size here. Forty timings taken at 11:30 on day two will give you a comfortable mean that describes a door nobody is queueing at. Take the sample during the twenty minutes either side of your busiest bin, take it at the lane in the middle of the bank rather than the one nearest the wall, and time every consecutive person including the ones who fumble. The instinct to skip the difficult cases is exactly what produces a service time estimate that predicts no queue at a door that is visibly queueing.

Door capacity is lanes divided by mean service time

The arithmetic is simple enough to do in a meeting, which is the point of doing it.

One lane at 4.2 seconds passes 3,600 divided by 4.2, which is 857.1 people per hour. Eight lanes pass 6,857 per hour.

Cut the mean service time to 3.1 seconds and one lane passes 1,161.3 per hour, so eight lanes pass 9,290. The same hardware, the same badges, the same staff count, and 2,433 more people an hour, a lift of 35.5 per cent.

Where does the 1.1 seconds come from? Almost all of it from recovery. Halve the first-tap failure rate and the recovery component falls from 1.2 seconds to 0.6. Move badge collection so people are holding a badge before they reach the lane and the approach component falls by half a second. Neither change involves buying anything.

The relationship is nonlinear in the direction that hurts. Going from 4.2 to 3.1 seconds is a 26 per cent cut in service time and a 35.5 per cent gain in capacity, because capacity is the reciprocal. Going the other way is just as sharp, so a door that degrades from 4.2 to 5.0 seconds because it started raining and everybody is wearing coats loses 16 per cent of its capacity precisely when the arrivals are worst.

What does capacity look like against a real arrival rate?

Capacity on its own settles nothing. Put it against arrivals and the picture becomes operational.

Say 9,000 people arrive in the peak hour, which is 2.5 per second. Eight lanes at 4.2 seconds deliver 1.905 per second. The deficit is 0.595 per second, which is 35.7 people per minute joining a queue that never clears. Across a 30 minute peak, 1,071 people accumulate outside the door, and the queue keeps growing until the arrival rate falls.

The same eight lanes at 3.1 seconds deliver 2.581 per second against 2.5 arriving. The queue clears, slowly, at utilisation of 96.9 per cent. Anyone who has worked a queue knows what 97 per cent utilisation feels like: it works until one badge jams, and then it does not.

Ten lanes at 3.1 seconds deliver 3.226 per second, which is 11,613 per hour and utilisation of 77.5 per cent. That is a door with enough margin to absorb a failed reader without a visible queue.

Little published the general result in Operations Research in 1961: the mean number in a system equals the arrival rate multiplied by the mean time each unit spends in it. At 2.5 arrivals per second, every additional second of average waiting puts another 2.5 people in the queue area. Ninety seconds of average wait means 225 people standing outside, which is a floor space and stewarding question before it is a data question.

The shape of those arrivals across the morning is a separate piece of work, and it is the one that decides how many lanes you open at 09:00 against 11:00. That belongs with the arrival curve.

Where the throughput number gets misread

Three misreadings are common enough to name.

The first treats mean service time as if the variation did not matter. A lane with a mean of 3.1 seconds and a long right tail behaves worse than a lane with a mean of 3.4 seconds and no tail, because the tail events block everybody behind them. Record the distribution when you run your stopwatch sample, and look at the 90th percentile as well as the mean.

The second treats all lanes as equal. The lane nearest the door gets more traffic, the lane at the far end gets less, and a mean across eight lanes hides a first lane running at 120 per cent of its fair share. Signage and a steward directing people fixes more capacity than another reader would.

The third counts the wrong event. If your throughput figure comes from grant events in the panel log, it is counting successful reads, and every failed first attempt is either invisible or is inflating the count depending on how the panel logs it. Knowing which of those is happening means reading the access control event stream properly, including the denies.

That third one deserves a moment, because it flatters you in both directions at once. A panel that logs only successful grants makes your first-tap failure rate look like zero, so the recovery component of service time disappears from the model while continuing to exist in the queue. A panel that logs every attempt as a separate row makes your throughput look higher than the number of people who actually got in. Neither version is usable until you know which one you have, and the fastest way to find out is to stand at the lane for five minutes with the log open beside you.

Where this stops

Capacity arithmetic assumes the constraint is the lane. Often it is not.

At many venues the real constraint sits upstream, in the walkway from the station, the escalator, or the security screening a venue operates in front of your readers. A door with 11,613 people per hour of scanning capacity behind a screening line that passes 4,000 has 4,000 of usable capacity, and the money spent on the ninth and tenth lane bought nothing.

The other limit is that the mean service time you measured is a property of that door, that crowd and that weather. It moves between days, it moves between shows in the same hall, and it moves when you change badge stock. Treat it as something to re-measure each edition rather than a constant to look up.

And the measurement itself needs the same scepticism you would apply to any sensor reading, which is the argument in entrance counting accuracy. A stopwatch operated by a tired person at 09:10 has its own error, and two observers timing the same lane will not agree exactly. Both of those belong in the same attendee analytics discipline as the counts themselves.

Start this week by timing forty people at your busiest lane at the busiest moment of your next show, writing down presentation to gate release for each. Divide the lane count by the mean and compare the answer with the number of people who actually arrived in that hour. The gap between those two figures is the queue you have been staffing around.