wireless communication

TPUNB at 100 nodes: polling math the spec sheet skips

A 100-terminal rating says nothing about round time. Airtime, turnaround and slotted-ALOHA math for 100 nodes, and the polling interval each symbol rate holds.

A gateway datasheet answers one question about capacity: 100 terminals. That number goes into the bill of materials, the network diagram and the tender response, and it is almost always read as a promise about speed. It is not. It says nothing about how long a full round takes, nothing about what happens when two nodes answer at once, and nothing about the margin you have left when a meter misses its first slot and needs a second one.

In a real deployment the question is never "how many nodes". It is "can this network finish a complete round inside my polling interval, every interval, with room for retries". For a water utility that interval is often five minutes; for a substation it can be ten seconds. The same gateway answers both requests very differently, and the arithmetic that separates them fits on one page.

This is the calculation behind the TPUNB platform our terminals run on, worked through for 100 nodes at each of the three symbol rates the hardware supports. Every assumption is labelled, so you can substitute your own frame sizes and redo it.

What "100 terminals" promises, and what it does not

TPUNB is a narrowband S-FSK system with a TDMA link layer that also carries a slotted-ALOHA mode. The vendor describes three resource allocation strategies: deterministic pre-allocation, ALOHA random scheduling and dynamic scheduling, with a minimum network latency of 0.3 s. The same page gives a number that is much more useful for sizing: one gateway handles 40 terminals concurrently within 2 seconds at a 99% success rate (techphant.cn).

Divide one by the other. Two seconds across 40 terminals is 50 ms per terminal, including the downlink that tells it to speak and the uplink that carries the reading. That 50 ms figure is the whole article in one number: it is the per-terminal budget the vendor's own claimed throughput implies, and everything below is an attempt to derive it from first principles and then scale it to 100.

If you accept 50 ms per terminal, 100 terminals is a five-second round. Whether five seconds is good or catastrophic depends entirely on your polling interval — and on whether those five seconds are spent transmitting or waiting for a radio to turn around.

The four numbers that decide the round

A polled transaction has four components, and only two of them scale with the radio's bit rate.

  1. Downlink poll. The gateway addresses one terminal. I assume a 6-byte payload — address plus command.
  2. Turnaround. The terminal wakes, switches from receive to transmit, and starts its preamble. I assume 15 ms, which is the sort of figure a low-power node with an RC oscillator and a cold PLL produces; it should be measured, not trusted.
  3. Uplink reply. I assume a 12-byte payload: 4-byte meter address, 4-byte reading, 2-byte status, 2-byte application checksum.
  4. Second turnaround. The gateway goes back to receive, or the next poll starts.

Airtime itself is simple: bytes × 8 / symbol_rate. What is not obvious is how many bytes are on the air before your payload starts. I assume 14 bytes of framing — an 8-byte preamble and sync word, a 4-byte PHY header and a 2-byte CRC. That is an assumption, not a spec sheet value, and it is the single number in this model most worth checking against your own captures.

Symbol rate

Poll frame (20 B)

Reply frame (26 B)

2 × turnaround

Full transaction

100 terminals

2.4 kbps

66.7 ms

86.7 ms

30 ms

183 ms

18.3 s

19.2 kbps

8.3 ms

10.8 ms

30 ms

49 ms

4.9 s

76.8 kbps

2.1 ms

2.7 ms

30 ms

35 ms

3.5 s

Time to poll every terminal once, from 5 to 150 nodes, at the three symbol rates the TP5510 and TP5600 support

The model predicts 40 terminals in 1.97 s at 19.2 kbps. The vendor claims 40 terminals in 2 s. That agreement is the reason to trust the 100-node column: nothing was tuned to make it land there, and the same arithmetic then says a full round at the slowest rate costs 18.3 seconds.

Turnaround is the floor nobody puts on the datasheet

Look at what happens between 19.2 and 76.8 kbps. The bit rate rises by 4×, the airtime falls by 4× — and the transaction time only drops from 49 ms to 35 ms, a 29% improvement. At the fast rate, 30 of those 35 milliseconds are turnaround.

Per-transaction budget at each symbol rate; the grey blocks are radio turnaround, which does not scale with bit rate

At 2.4 kbps turnaround is 16% of the slot; at 76.8 kbps it is 86%. This has two consequences that are easy to get wrong in a design review:

  • Raising the symbol rate has diminishing returns. Going from 2.4 to 76.8 kbps — 32× the raw rate — buys a 5.3× shorter round, not 32×.
  • A slow rate is not as expensive as it looks. Most of the airtime at 2.4 kbps is spent on payload, which is exactly what a long-range link needs. If your frames are small, the fast rates are mostly buying you turnaround.

If you need a shorter round than 35 ms per terminal, the next lever is not the radio. It is the protocol: skip the downlink poll entirely and let terminals transmit in pre-assigned slots, halving the transaction by removing one frame and one turnaround.

The hardware at each end

The two products in this calculation sit at opposite ends of the same link and are usually specified separately, which is how mismatched symbol rates get ordered.

Item

TP5510 gateway

TP5600 terminal

Role

Concentrator, up to 100 terminals

Node / transparent data terminal

Frequency

470–510 MHz

470–510 MHz

Channel bandwidth

≤200 kHz per carrier

≤200 kHz per carrier

Symbol rate

2.4 / 19.2 / 76.8 kbps

2.4 / 19.2 / 76.8 kbps

Transmit power

26 ± 1 dBm

17 dBm, configurable

Sensitivity

−116 dBm @19.2 kbps

not stated in the spec

Channel count

1

1

Modes

Broadcast and directed (unicast)

Bidirectional transparent, unidirectional RX / TX

Interfaces

RS485 ×1, RS232 ×1

RS485 ×1, RS232 ×1

Supply

DC 12 V (5–28 V), ≤1 W, ~0.25 W average

DC 12 V (8–28 V), ≤1 W

Enclosure

IP65, outdoor, −40 to +70 °C

Indoor, −40 to +70 °C

Both ends support all three rates, so the rate axis in the tables above is a configuration decision, not a hardware limit. That is worth stating plainly because it is also the most common field fault: a gateway left at 2.4 kbps while the terminal is set to 19.2 kbps, or the reverse. The link still works at short range, the round is eight times longer than designed, and nobody notices until the node count grows.

Scheduled slots and slotted ALOHA are different networks

Everything above assumes a scheduler that hands out slots. TPUNB also runs a slotted-ALOHA mode, and the two behave nothing alike once the node count passes a few dozen. Full specifications are in the TP5510 product page and the TP5600 page; the behaviour is generic to any random-access scheme.

In slotted ALOHA, throughput follows S = G·e^−G, which peaks at 1/e ≈ 0.368 when the offered load G equals one frame per slot. Past that peak, adding traffic reduces the number of successful frames. That ceiling has been known since the ALOHANET work and it is still the reference model for LPWAN random access (arXiv:1611.05569). Worse, the peak is not sustainable if you also care about delay: with binary exponential backoff, the throughput that keeps the mean queue bounded is about 0.216, not 0.368 (arXiv:0801.4054).

At 2.4 kbps a 26-byte reply occupies 86.7 ms, so the channel offers 11.5 slots per second. Put 100 nodes on it with no scheduler and vary the reporting interval:

Reporting interval

Offered load G

Attempts that survive

Note

900 s

0.010

99.0%

comfortable, but slow to notice a dead node

300 s

0.029

97.2%

the usual meter-reading setting

60 s

0.144

86.5%

13.5% of attempts collide and need a retry

30 s

0.289

74.9%

a quarter of the traffic is now retransmission

10 s

0.867

42.0%

near the throughput peak, and unstable in practice

Slotted-ALOHA throughput curve with the operating points for 100 nodes at five reporting intervals

The scheduled mode does not have this curve. A pre-allocated slot is exclusive, so throughput stays linear in offered load until the frame is full — the diagonal line in the figure — and the 18.3 s round at 2.4 kbps is a hard, repeatable number rather than a distribution. The practical split is the obvious one: scheduled slots for the routine reading round, ALOHA for alarms, join requests and anything that cannot wait for its slot. Mixing them by accident is the failure mode below.

Failure modes that quietly break the arithmetic

Everyone defaults to ALOHA. A network commissioned with the default random-access mode looks fine at 30 nodes and degrades at 100. The symptom is not a dead network, it is a round that takes 20% longer every month and a growing retry count in the gateway log.

Retries are charged at full airtime. Each retry costs another 183 ms at 2.4 kbps. If 1% of 100 nodes need one retry per round, that is 183 ms — invisible. If a noisy meter room pushes the loss rate to 10%, you have added 1.8 s to every round, and the load that produced the loss is now also producing the retransmissions.

Clock drift eats the guard band. Slotted operation only works if nodes agree on when a slot starts. Published slotted-ALOHA overlays on low-cost hardware report 5.37 ms of synchronisation accuracy with better than 99% success (arXiv:1902.09468). Against an 86.7 ms frame at 2.4 kbps that is 6% of a slot; against a 2.7 ms frame at 76.8 kbps it is twice the frame. Fast symbol rates and loose timing do not mix.

Half-duplex collisions with the poll. The gateway cannot hear a reply while it is transmitting the next poll. Any node that answers late — because it woke late, or because its reply was queued behind a serial read — collides with the gateway's own downlink. This is why the turnaround figure matters twice: it is both lost time and the window in which a late node ruins somebody else's slot.

A firmware update takes the whole channel. The platform claims a differential update in under three minutes. Three minutes of broadcast traffic is 36 polling rounds at 2.4 kbps. Schedule updates outside the billing window, and treat the update as a capacity event rather than an operations detail.

A worked example: 100 water meters on a five-minute cycle

Take the common case: 100 meters, one reading every five minutes, TP5510 at the top of the riser, TP5600 terminals in the meter cabinets, 12-byte readings.

At 2.4 kbps the round is 18.3 s, which is 6.1% of a 300-second interval. That leaves a 16× margin before the channel is full — enough for a 10% retry rate, a handful of alarm frames and a join storm after a power cut. At 19.2 kbps the same round is 4.9 s, or 1.6%, and you could run a 60-second cycle at 8.2% occupancy.

The numbers flip if you keep 2.4 kbps and ask for 60-second readings: 30.6% occupancy is survivable but leaves almost nothing for retries, and the moment 10% of nodes need a second attempt you are at 33%. That is the case where the sales sheet says 100 terminals, the arithmetic says 100 terminals, and the network still feels slow — because the interval, not the capacity, was the constraint.

A rule of thumb that has held up in our deployments: keep scheduled occupancy under 30% of the interval and reserve the rest for retries, alarms and firmware. If the arithmetic says you are above 30%, either extend the interval, move to 19.2 kbps, or split the site across two gateways. Adding nodes does not help, and neither does raising transmit power.

How to verify the round time on site

The model is only useful if you can check it against the real network. Six measurements, all from the gateway's own log:

  1. Timestamp every poll and every reply. Inter-arrival time per terminal should cluster around the transaction time in the table above; a bimodal distribution means retries.
  2. Sum the round. Time from the first poll of a round to the last reply, divided by the node count, gives the realised per-terminal cost. More than 20% above the model means turnaround, slot padding or retransmission.
  3. Count retries, not failures. A network with a 2% retry rate and a 99% delivery rate is working harder than one with 0.2% and the same delivery.
  4. Read back both symbol rate settings. The gateway and the terminal must agree; the link is silent about mismatches until you measure the round.
  5. Check the tail, not the average. The p95 per-terminal time is what breaks a polling deadline. An average of 50 ms with a p95 of 400 ms is a network that misses its window several times a day.
  6. Log the update window. If a firmware campaign overlaps the polling round, the round time doubles for three minutes and the alarm history looks like a radio fault.

A minimal check you can run against an exported log, assuming one line per transaction with a millisecond timestamp:

nodes, rate_kbps = 100, 19.2
poll_ms  = (6 + 14) * 8 / rate_kbps
reply_ms = (12 + 14) * 8 / rate_kbps
tx_ms    = poll_ms + 15 + reply_ms + 15      # 15 ms turnaround, measured
print(f"round {nodes * tx_ms / 1000:.2f} s at {rate_kbps} kbps")
print(f"occupancy at 300 s: {nodes * tx_ms / 300000:.1%}")

Run it with your own payload sizes and your own measured turnaround. If the printed occupancy is under 30% and the log agrees within 20%, the network has margin. If the log disagrees with the model, believe the log — and find out which of the four numbers is wrong.

FAQ

Can one gateway really serve 100 terminals?

Yes, for a five-minute reading cycle at 2.4 kbps: 100 terminals cost 18.3 s per round, 6.1% of the interval. The same gateway cannot give 100 terminals a reading every ten seconds — at 10 s the round would occupy 183% of the interval, which is another way of saying the scheduler would simply fall behind.

Does a faster symbol rate always shorten the round?

No. Turnaround is a fixed 30 ms per transaction in this model, so the gain from 2.4 to 76.8 kbps is 5.3×, not 32×. At 76.8 kbps, 86% of each slot is turnaround. If you need more speed than that, remove the downlink poll and use pre-assigned uplink slots.

What happens if I leave 100 nodes in ALOHA mode?

At a 60-second reporting interval the offered load is G ≈ 0.14 and about 13.5% of attempts collide; at 30 s it is 25%. The network still delivers, but a growing share of its airtime is retransmission, and the latency tail grows faster than the average.

How much margin should I leave?

Keep scheduled occupancy at or below 30% of the polling interval. The remaining 70% pays for retries, alarm frames, join requests and firmware updates. Above 30%, the first bad week — a storm, a firmware campaign, a batch of failing meters — turns into a missed polling window.

If you are sizing a network now, the two questions worth answering before the node count are the polling interval and the payload size; the product catalogue lists the gateway and terminal combinations that support all three symbol rates, and the earlier HaLow versus sub-GHz comparison covers the case where the payload is a video frame rather than a meter reading.

More in wireless communication