How Can One-Way Delay Be Modeled for 5G Low Latency Communications?

Queueing delay dominates one-way delay at high load

The paper starts from a practical problem: delay-based congestion control needs to know when queues begin to build at the bottleneck, but one-way delay (OWD) is hard to measure because sender and receiver clocks are not synchronized. Round-trip time has its own flaw, since queueing on the reverse path distorts the signal. The authors instead ask whether the queueing delay (QD) at the base station can be reconstructed from components that are easier to observe.

Their answer rests on where the bottleneck sits. In the edge architecture they study, the edge server is close to the base station and the wireless link is slower than the wired backhaul, so the last buffer in the wired domain, the radio link control (RLC) buffer at the base station, is where packets accumulate. The paper therefore models OWD as QD plus a term it calls Tsys, which covers processing time, transmission time interval and alignment delay that are fixed or bounded by 5G specifications.

The simulations support treating QD as the dominant part, but only conditionally. Near network saturation, where resource utilization approaches its maximum, QD rises above 90% of OWD in both the single-UE and multi-UE scenarios. At low loads the share of QD in OWD actually falls slightly, because OWD grows faster than QD. That split matters: the approximation is strongest exactly in the high-rate, near-saturation region the paper identifies as the operating range for low-latency applications.

Aggregation delay and channel access time split the queue

The paper's central move is to decompose QD for a packet carried in a transport block (TB), the unit the MAC layer hands to the physical layer for one over-the-air transmission. Packets are grouped into a TB, and the number grouped is the aggregation level. Because all packets in one TB leave the RLC queue together, they share a departure time but not an arrival time, and that difference is the first component.

Aggregation delay (AD) is how long a packet waits in the RLC queue until the last packet of its TB arrives. The first packet in a TB therefore has the largest AD and the last packet has zero. The second component, channel access time (Taccess), is the interval between the arrival of that last packet and the departure of the TB from the queue. QD for a packet is then AD plus Taccess.

The decomposition is useful because the two parts are observable in different places. The edge server stamps each packet with its send time, and the UE can read that timestamp from the payload, so AD can be measured at the UE without synchronized clocks. Taccess cannot be measured the same way, but the paper argues it can be estimated from network conditions, and for the last packet in a TB it equals that packet's QD, which links the estimate back to an observable OWD.

The UE measures aggregation delay from MAC timestamps

The measurement path depends on a specific property of the radio link: packets aggregated into one TB arrive at the MAC layer of the UE at the same time and carry the same MAC timestamps. A UE with access to MAC layer information can therefore identify which packets were grouped together and read the aggregation level directly. The paper names open-source UE stacks and monitoring tools as ways to obtain this, running on commercial handsets.

With that grouping known, AD for each packet follows from the send timestamps the edge server placed in the payload. The paper justifies replacing arrival times at the RLC queue with send times at the edge server by assuming an ideal backhaul, consistent transmission time from server to base station along the same path, and no reordering in the physical link. Under those assumptions the difference in send times equals the difference in queue arrival times.

This is where the model's scope is narrowest. It assumes all traffic toward a UE shares a single data radio bearer and thus one RLC queue, that packets fit within one maximum transmission unit without segmentation, and that the RLC layer runs in unacknowledged mode, which does not retransmit. The authors note the approach extends to non-identical flows because different bearers use dedicated buffers, but the single-queue assumption is what makes the per-packet arithmetic clean.

Channel access time separates drained queues from backlog

Taccess is the harder half, and the paper analyzes it by splitting network operation into two scenarios after the last packet of a TB reaches the RLC queue. In the first, the MAC scheduler grants enough physical resource blocks at the next scheduling decision to send the TB, so Taccess stays between zero and one scheduling interval. If a run of consecutive TBs stays in that range, the queue is being drained each scheduling round without a persistent backlog.

In the second scenario the TB is not granted transmission and must wait for later scheduling periods, spaced by whole scheduling intervals. This is the regime where queueing builds, and the paper's simulation results show Taccess becoming the dominant part of QD there, exceeding 95% once the network congests. Before congestion, the balance is reversed: AD reaches peaks of 80% of QD in the single-UE case and 70% in the multi-UE case near maximum resource utilization, while Taccess falls to 20% and 30% respectively.

The two scenarios give the model its practical reading. A Taccess bounded within one scheduling interval signals a drained queue, while Taccess stretching across multiple scheduling periods signals backlog. Because the scheduler, cross-traffic, buffer sizes and channel quality all feed into Taccess, the paper treats it as something to estimate from conditions rather than measure directly.

Simulated delay curves match the two-part sum

The validation uses NS3 with the 5G-LENA module, a per-UE queueing policy at the base station, a line-of-sight indoor office channel over 40 MHz in the n78 band, numerology 0 with 15 kHz subcarrier spacing, a round-robin scheduler, fixed MCS 28, 10 dBm transmit power, UDP transport and a 1500-byte MTU. Runs cover single-UE and two-UE cases, the latter pairing a UDP flow with a TCP Cubic flow as cross-traffic, across low and high send rates.

In the reported curves, the sum of AD and Taccess tracks the QD curve across loads and scenarios. Two features stand out. Within a TB, the first packet shows the highest delay and the last the lowest, matching the aggregation logic. And OWD and QD follow similar sawtooth patterns caused by grouping multiple packets into one TB, with a persistent gap of about 2.5 ms between them that the paper attributes to constant or bounded latency sources such as processing time and the transmission time interval.

That gap is the empirical face of Tsys. The paper notes Tsys can be comparable to QD under low network loads, which is another reason the QD-dominates-OWD claim is load-dependent rather than universal. The multi-UE case also shows a higher QD share of OWD than the single-UE case, which the authors attribute to larger Taccess from scheduling latency when another user competes for resources.

Simulation evidence leaves field validation open

All reported validation comes from simulation, not from a deployed network. The authors state that they plan to evaluate the model experimentally on open-source 5G platforms across different scenarios and traffic types, and to use the framework to design an end-to-end rate allocation policy at the transport layer. Until then, the accuracy claims rest on the NS3 configuration described, including its ideal backhaul and unacknowledged-mode assumptions.

The model also inherits limits from what it chooses to model. It targets the RLC buffer as the single bottleneck and folds everything else into a bounded Tsys, so a path where latency between edge server and base station is significant would need that latency added into Tsys, as the paper itself notes. The claim that the framework accounts for retransmissions sits alongside the unacknowledged-mode assumption used in the analysis, so readers should treat the retransmission case as asserted rather than demonstrated in the results shown.

Bound channel access time before trimming aggregation

The paper's most decision-relevant result is a priority ordering rather than a single number. Before congestion, aggregation delay can account for most of QD, peaking at 80% in the single-UE scenario and 70% with cross-traffic. Once resource utilization climbs and the queue backs up, channel access time takes over at more than 95% of QD. The lever therefore depends on which regime the link is in.

For a link near saturation, the model points to bounding Taccess as the binding constraint, with aggregation level monitored to regulate AD. That corresponds to the paper's first scenario, where Taccess stays within one scheduling interval and the queue drains each round. The practical threshold the framework offers is that boundary: a Taccess that fits inside a single scheduling interval indicates no persistent backlog, while one spanning multiple scheduling periods marks the onset the paper wants congestion control to detect.

The value of the decomposition is that it turns an unmeasurable quantity at the base station into two parts with different observability, one read at the UE from MAC timestamps and one estimated from network conditions. Whether that substitution holds in a real 5G deployment, rather than in the simulated edge architecture with its ideal backhaul, is the question the authors leave to future experimental work.

Sources used in this answer