HARQ over a Long Round-Trip in 5G NTN
Why long RTT starves stall-and-wait HARQ, the move to up to 32 processes, enabling/disabling HARQ feedback per process, and relying on RLC ARQ for GEO.
NR HARQ is stop-and-wait: each HARQ process holds one transport block and sits idle until its ACK/NACK returns. On the ground that idle gap is tiny. Over a satellite the round-trip is tens to hundreds of milliseconds, so a handful of processes would spend almost all their time waiting — the pipe empties and throughput collapses. NTN fixes this two ways: many more HARQ processes, and the option to switch HARQ feedback off per process and lean on RLC instead.
Introduction
Hybrid ARQ (HARQ) is NR's fast, physical-layer error-recovery mechanism: the receiver combines retransmissions with the original transmission (soft combining) and asks for another copy only if it still cannot decode. It is organised as a set of parallel stop-and-wait processes — each process handles one transport block at a time and cannot be reused until that block's fate (ACK, or NACK plus retransmission) is settled. On a terrestrial link the feedback comes back in a slot or two, so the processes cycle quickly and the mechanism is nearly free.
NTN breaks that assumption in the most basic way possible: it makes the round-trip enormous. A GEO path is roughly 540 ms there-and-back; even a LEO path is single-digit to low-tens of milliseconds. Stop-and-wait over a delay that large means each process spends the overwhelming majority of its life idle, waiting for a receipt that is still in flight — and with a terrestrial-sized pool of processes the transmitter runs out of processes to use and the link simply stalls.
This page is about the two levers NTN pulls to keep the link full despite that delay: raising the HARQ process ceiling so many more transport blocks can be outstanding at once, and allowing HARQ feedback to be switched off per process so the transmitter need not wait at all, moving reliability up to RLC. You meet these mechanisms whenever you dimension throughput or latency on a satellite bearer — they are the reason an NTN link can approach its capacity instead of idling away most of it.
On this page
Why a long round-trip breaks stop-and-wait
In plain words: imagine posting letters that each need a reply before you send the next one. Across town the reply comes back the same day, so a couple of letters keeps a steady conversation going. Now post to the far side of the world: each reply takes weeks, so if you insist on waiting for one before sending the next, you send almost nothing. Two fixes exist — keep many letters in the post at once (more HARQ processes), or stop demanding a reply on routine letters and reconcile any that go missing later (disable feedback, lean on RLC). NTN uses both.
Recall the baseline (see HARQ): a process transmits a TB, then cannot be reused until feedback comes back and any retransmission is resolved. The number of processes sets how many transport blocks can be "in flight" at once. Fill time is roughly (number of processes) × (TB airtime); if that product is smaller than the round-trip, every process is stalled waiting and the link idles.
Put numbers on it. If a transport block occupies one slot of ~0.5 ms (30 kHz SCS) and the GEO round-trip is ~540 ms, keeping the pipe full would need on the order of 540 / 0.5 ≈ a thousand outstanding blocks — far beyond any HARQ process count. Even doubling the ceiling cannot fully cover GEO, which is exactly why GEO needs the second lever (disabling feedback) rather than more processes alone. For LEO, where the round-trip is a few milliseconds, the arithmetic lands within reach of the extended process count, so raising the ceiling is often enough on its own.
HARQ is per-process stop-and-wait in the MAC (TS 38.321): a process is busy from the moment it transmits until its feedback (and any retransmission) completes.
With a GEO round-trip near 540 ms, a terrestrial-sized set of processes drains in a few slots and then waits ~half a second doing nothing. The pipe is mostly empty, so the achievable rate is a fraction of the link capacity.
Two levers: (1) allow more processes so more TBs are in flight to fill the long pipe, and (2) let the network disable feedback on a process so it need not wait at all, moving reliability to RLC ARQ.
The mental model: think of the link as a long pipe. To keep it full you either send more parcels before the first receipt returns (more HARQ processes), or you stop waiting for receipts entirely (disable feedback) and reconcile losses later at a higher layer (RLC AM).
Fix #1 — keep the pipe full with up to 32 processes
Terrestrial NR allows up to 16 DL HARQ processes. NTN doubles the ceiling to up to 32 DL HARQ processes, configured through the extended nrofHARQ-ProcessesForPDSCH. More processes means more transport blocks can be outstanding at once, so the transmitter can keep launching new TBs across the whole round-trip instead of stalling.
Spec anchor: the extended count (up to 32 DL HARQ processes) and its configuration via nrofHARQ-ProcessesForPDSCH are in TS 38.331 / TS 38.321; the timing relationships are in TS 38.213.
Fix #2 — disable HARQ feedback per process
Adding processes helps, but there is a deeper problem for GEO: even if you keep the pipe full, a HARQ retransmission triggered by a NACK arrives roughly a full round-trip late. For GEO that is over half a second — far too late to matter for most traffic, and the soft buffer for that process was tied up the whole time. So NTN allows the network to enable or disable HARQ feedback per HARQ process, via a per-process flag (downlinkHARQ-FeedbackDisabled).
A per-process configuration marking each DL HARQ process as feedback-enabled or feedback-disabled. When disabled, the UE sends no HARQ-ACK for TBs on that process.
For GEO (round-trip ~540 ms) a retransmission would arrive far too late to help interactive or latency-bound traffic, and waiting ties up the process and its soft buffer. Better to schedule blindly and free the process immediately.
With feedback disabled the gNB does not wait for ACK/NACK — it schedules the process again right away, and reliability is provided by RLC AM ARQ (plus conservative link adaptation), not by fast HARQ retransmission.
The choice is orbit-dependent. For GEO, DL HARQ feedback is typically disabled because retransmissions cannot arrive in time to be useful. For LEO, where the round-trip is much shorter (single-digit to low-tens of ms), feedback may stay enabled so HARQ still recovers errors quickly. Because the flag is per-process, a deployment can even mix modes — some processes feedback-enabled for signalling, others disabled for bulk data.
What provides reliability when HARQ feedback is off
Turn off fast HARQ retransmission and you must recover robustness elsewhere. NTN leans on three things: conservative MCS and/or repetition so the first transmission is more likely to decode without help; and RLC AM ARQ, which detects missing PDUs from status reports and retransmits at Layer 2 — slower than HARQ but reliable. Because that safety net now runs at RLC/PDCP timescales stretched by the delay, the reassembly and discard timers must be scaled: RLC t-Reassembly and the PDCP discard timer are enlarged so the long round-trip does not trip them prematurely (see NTN timers). The extra coding margin ties back to low-SNR coverage design in NTN spectrum.
In plain words: with HARQ feedback disabled the network stops asking "did you get it?" on that process. It instead sends each block robustly enough to usually land first time, and relies on RLC AM to clean up the rare miss — trading HARQ's speed for RLC's guaranteed delivery.
HARQ and DRX timing scaled for the delay
Whether or not feedback is on, the timing constants around HARQ must reflect the round-trip. The HARQ round-trip time and the DRX retransmission-related timers (for example drx-HARQ-RTT-TimerDL/drx-HARQ-RTT-TimerUL and the associated retransmission timers) are extended so the UE does not start monitoring for a retransmission before it could physically arrive; K_offset participates in these relationships too (see NTN timers).
Uplink HARQ works the same way in NTN: UL transmissions are scheduled, and there is no explicit ACK channel to the UE — a new grant with toggled NDI means "received," a grant with the same NDI is an implicit retransmission request. When UL HARQ feedback is effectively disabled for a process, there is no HARQ-based retransmission at all and reliability again rests on RLC ARQ.
TN ↔ NTN: terrestrial NR runs at most 16 DL HARQ processes, always with feedback active, and its DRX HARQ-RTT and retransmission timers are slot-scale because the round-trip is a slot or two. NTN keeps the identical HARQ machinery but re-parameterises all three: up to 32 processes, per-process feedback that can be switched off, and RTT-scaled timers that also fold in K_offset. Nothing about the protocol is new — the same stop-and-wait, the same soft combining, the same NDI toggle — only the numbers are stretched to fit a satellite delay.
| Aspect | Terrestrial NR | NTN |
|---|---|---|
| Max DL HARQ processes | 16 | up to 32 (nrofHARQ-ProcessesForPDSCH extended) |
| HARQ feedback | Always active per process | Per-process enable/disable (downlinkHARQ-FeedbackDisabled) |
| Typical GEO choice | — | DL feedback disabled (retransmission too late) |
| Typical LEO choice | — | Feedback may stay enabled (short RTT) |
| Reliability fallback | HARQ, then RLC AM | Conservative MCS / repetition + RLC AM ARQ |
| Timers | Slot-scale | DRX HARQ-RTT / retransmission timers scaled by RTT + K_offset |
NTN HARQ parameters at a glance
The knobs the network turns to make HARQ survive the long round trip (TS 38.321 / TS 38.331 — see the ASN.1 for exact bounds):
| Parameter (IE) | Terrestrial NR | NTN | Why |
|---|---|---|---|
nrofHARQ-ProcessesForPDSCH | up to 16 | up to 32 | More transport blocks in flight to fill the long pipe. |
downlinkHARQ-FeedbackDisabled | — (always on) | per-process enable/disable (bitmap) | Disable where a retransmission would arrive too late (GEO); rely on RLC ARQ. |
drx-HARQ-RTT-TimerDL / ...UL | short | extended for RTT | DRX must not expect a (re)transmission before the round trip elapses. |
drx-RetransmissionTimerDL / ...UL | short | extended for RTT | Window for a possible retransmission widened to the satellite delay. |
| Reliability fallback | HARQ retransmission | RLC AM ARQ + robust MCS/repetition | When feedback is off, Layer 2 and coding carry reliability (see Timer scaling). |
⚠ Common pitfalls / gotchas
- Adding processes and expecting GEO to fill. Even 32 processes fall far short of the ~1000 outstanding blocks a 540 ms GEO round-trip would need at slot-scale airtime — GEO fundamentally needs feedback disabling, not just more processes.
- Disabling feedback without scaling RLC/PDCP timers. With HARQ off, RLC AM is the safety net; if
t-Reassemblyand the PDCP discard timer are left at terrestrial values the long delay trips them and packets are discarded before they can be recovered. - Leaving DRX HARQ-RTT timers slot-scale. The UE would wake to monitor for a retransmission before it could physically arrive, wasting power and potentially missing the real one.
- Applying GEO's feedback-off recipe to LEO. LEO's few-ms round-trip is short enough that HARQ still recovers errors usefully, so blanket-disabling feedback throws away fast recovery for no gain.
- Forgetting soft-buffer occupancy. A feedback-enabled process ties up its soft buffer for the whole round-trip; with a long delay and too few processes this becomes a hidden bottleneck independent of the process-count arithmetic.
Summary
HARQ over a satellite is the same stop-and-wait mechanism as on the ground, defeated by one number: the round-trip delay. Because a process is idle from transmit until its feedback returns, the number of processes caps how many transport blocks can be in flight, and when (processes × TB airtime) is smaller than the round-trip the link simply stalls. A GEO round-trip near 540 ms would need on the order of a thousand outstanding blocks to stay full — far beyond any process count.
NTN answers with two levers. First, raise the ceiling to up to 32 DL HARQ processes (nrofHARQ-ProcessesForPDSCH) so more blocks are outstanding — enough for LEO's short round-trip, though not for GEO. Second, allow HARQ feedback to be disabled per process (downlinkHARQ-FeedbackDisabled): the gNB schedules blindly and frees the process immediately, which is the typical GEO choice since a retransmission would arrive far too late anyway. LEO often keeps feedback enabled because its round-trip is short.
When fast HARQ retransmission is off, reliability moves up the stack: conservative MCS and repetition make the first transmission likely to decode, and RLC AM ARQ cleans up the rare miss at Layer 2. That only works if the surrounding timers — RLC t-Reassembly, the PDCP discard timer, and the DRX HARQ-RTT / retransmission timers — are all scaled to the satellite delay and K_offset. Get the process count, the feedback mode, and the timer scaling right for the orbit, and an NTN link approaches its capacity instead of idling most of it away.
Q. Why does a long round-trip hurt stop-and-wait HARQ?
A. Each process is idle from transmit until its feedback returns. The number of processes caps how many TBs are in flight; if (processes × TB airtime) is smaller than the round-trip, every process ends up waiting and the link idles. With a GEO round-trip near 540 ms, a terrestrial-sized process set drains in a few slots and then stalls for half a second, so throughput collapses.
Q. Why allow up to 32 HARQ processes?
A. To keep the long pipe full. Doubling the ceiling from 16 to 32 (via the extended nrofHARQ-ProcessesForPDSCH) lets twice as many transport blocks be outstanding at once, so the transmitter can keep launching new TBs across the round-trip instead of stalling while it waits for the first ACK.
Q. Why disable HARQ feedback for GEO, and what replaces it?
A. A retransmission triggered by a NACK would arrive about a full round-trip later — over half a second for GEO — far too late to help, while tying up the process and its soft buffer. So the network disables feedback on those processes (downlinkHARQ-FeedbackDisabled), schedules blindly, and relies on RLC AM ARQ plus conservative MCS/repetition for reliability.
Q. When HARQ feedback is off, what still provides reliability?
A. Robustness comes from a more conservative MCS and/or repetition so blocks usually decode first time, and from RLC AM ARQ, which detects missing PDUs via status reports and retransmits at Layer 2. The RLC t-Reassembly and PDCP discard timers are scaled up so the long delay does not trip them.
Related NTN topics
HARQ's behaviour over the satellite link is set by the delay and the timers around it — follow those threads next.