Multi-Link Operation (MLO)
- Multi-Link Operation (MLO) is an IEEE 802.11be mechanism that allows a multi-link device to use multiple radio interfaces concurrently via a unified MAC interface.
- It separates upper MAC functions from link-specific lower MAC operations, enabling independent channel access with dynamic traffic steering and per-flow management.
- Research indicates that while MLO can significantly improve throughput and reduce latency under light loads, it may introduce contention challenges in dense network environments.
Multi-Link Operation (MLO) is the IEEE 802.11be mechanism that allows an access point and station, when implemented as Multi-Link Devices (MLDs), to maintain a single association while operating over multiple links with coordinated upper-MAC control and link-specific lower-MAC contention. In the literature, MLO is presented as a foundational Wi-Fi 7 capability for increasing aggregate throughput, reducing latency, and improving reliability, but its realized behavior depends strongly on link concurrency mode, traffic-to-link assignment, contention conditions, and coexistence with neighboring networks (López-Raventós et al., 2022, Carrascosa-Zamacois et al., 2022).
1. Architectural model and management plane
MLO is built around the MLD abstraction: a single logical 802.11 device comprising multiple radio interfaces, each with its own PHY and link-specific MAC state, while exposing a unified MAC service interface upward. The common upper MAC (U-MAC) performs link-agnostic functions such as sequence number assignment, MSDU aggregation and de-aggregation, fragmentation and defragmentation, buffering, and common management procedures; the lower MAC (L-MAC) exists per link and maintains its own EDCA queues, backoff process, and control-frame handling (López-Raventós et al., 2022, Liu et al., 19 Mar 2026).
A recurring architectural point across the literature is that MLO does not collapse multiple channels into a single bonded channel. Each link preserves independent channel access, and the U-MAC arbitrates how traffic is distributed across those independently contending links. This enables per-packet or per-flow steering, cross-link reordering at the receiver, and dynamic TID-to-link mapping. The architecture also extends management and discovery through the Multi-Link Element (MLE), reduced-neighbor reporting, multi-link setup on a single link, and power-save mechanisms such as multi-link TIM and multi-link TWT (López-Raventós et al., 2022).
This split is operationally significant because many of MLO’s observed gains and anomalies arise precisely from the tension between a shared transmission buffer at the U-MAC and independent contention at the L-MAC. Papers that focus on delay, queueing, or cross-layer optimization consistently treat this coupling as the central design problem rather than a secondary implementation detail (Carrascosa-Zamacois et al., 2022, Liu et al., 19 Mar 2026).
2. Operating modes and MAC semantics
The literature distinguishes several closely related taxonomies for MLO. One axis separates asynchronous, independently contending multi-link access from synchronous access with tighter inter-link coordination. Another separates single-radio opportunistic access from multi-radio concurrent access. Under different naming conventions, these appear as EMLSR or MLSR for single-radio operation and EMLMR or MLMR for multi-radio operation (López-Raventós et al., 2022, Carrascosa-Zamacois et al., 2022).
In asynchronous or STR-oriented operation, each link can run its own contention and transmit independently, which maximizes spectrum-opportunity exploitation. In synchronous or NSTR-oriented operation, inter-link timing is constrained, typically through end-time alignment or defer-transmission, to avoid intra-device coexistence interference and ACK overlap. Practical STR operation additionally requires sufficient inter-channel frequency separation and/or self-interference suppression; one cited example uses four 80 MHz channels with two in 5 GHz and two in 6 GHz and at least 160 MHz separation (Carrascosa-Zamacois et al., 2022, Jung et al., 1 Sep 2025).
| Mode family | Radio structure | Characteristic constraint or opportunity |
|---|---|---|
| EMLSR / MLSR | One full radio plus sensing-capable auxiliaries | Opportunistic link selection, but one transmission at a time |
| NSTR EMLMR / NSTR MLO | Multiple radios | Start-time and end-time alignment to avoid intra-device interference |
| STR EMLMR / MLMR | Multiple radios | Independent, asynchronous, concurrent transmissions across links |
Single-radio MLO is therefore not simply a reduced-throughput version of multi-radio MLO. It changes the access semantics by forbidding concurrent occupancy of multiple links, which can reduce starvation and improve fairness in contested environments. Multi-radio MLO, by contrast, exposes the full benefit of parallel backoff and concurrent service, but also introduces new interaction paths between links, queues, and neighboring BSSs (Carrascosa-Zamacois et al., 2022, Carrascosa-Zamacois et al., 2022).
3. Throughput, latency, and statistical multiplexing gains
In contention-free or lightly contended settings, the dominant MLO mechanism is parallelism in channel access. When one link is busy, an MLD can initiate or continue contention on another link rather than waiting behind a single-link transmission, which reduces access waiting, buffer buildup, and aggregation size. In one downlink-only simulation campaign with Poisson arrivals, 12,000-byte packets, 80 MHz channels, two spatial streams, and MCS 256-QAM 3/4, the offered load compatible with a median delay target of 1 ms increased from approximately 0.5 Gbps in single-link operation to approximately 1 Gbps with STR EMLMR:2 and approximately 2 Gbps with STR EMLMR:4 (Carrascosa-Zamacois et al., 2022).
The same study reported that at 0.5 Gbps offered load, the 99th-percentile delay was approximately 1.6 ms for single-link operation, approximately 0.7 ms for STR EMLMR:2, and approximately 0.5 ms for STR EMLMR:4. At 1 Gbps, single-link delay became unbounded because the offered load exceeded single-channel capacity, whereas STR EMLMR:2 remained at or below approximately 1.7 ms at the 99th percentile and STR EMLMR:4 at or below approximately 0.7 ms (Carrascosa-Zamacois et al., 2022).
A separate finite-load analysis modeled STR-MLMR as an queue and showed that MLO contains the $95$th-percentile delay as traffic increases, primarily because the effective backoff seen by a packet becomes the minimum of several backoff processes when multiple interfaces are idle. That analysis also found rapidly diminishing gains from adding extra interfaces beyond the first few, because once the queueing probability becomes small, latency is dominated more by service time and contention overhead than by server multiplicity (Bellalta et al., 2022).
Application-specific studies on augmented reality report the same qualitative pattern. With 80 MHz total bandwidth, a MHz MLO configuration supported up to 7 AR STAs under the Greedy traffic-to-link policy, versus 6 in single-link operation; with 160 MHz total bandwidth, MLO supported up to 10 STAs versus 8 in single-link operation. The bottleneck was the downlink video stream and the relevant KPI was whether the worst STA’s 99th-percentile delay remained within the packet delay boundary while maintaining packet success (Alsakati et al., 2023).
4. Pathologies, coexistence, and fairness limits
A central result of the MLO literature is that multi-link concurrency can harm delay when links are shared by multiple contending BSSs. In a crowded four-BSS scenario with symmetric load and four shared channels, STR EMLMR BSSs could occupy multiple links concurrently, temporarily blocking neighbors across all of their shared opportunities. At a total offered load of 2.5 Gbps across four BSSs, STR EMLMR:4 used two or more interfaces during approximately of active time, and a contending BSS found all four links occupied approximately of the time. Under the same condition, the 99th-percentile aggregated-packet count increased from approximately 207 in single-link operation to approximately 317 in STR EMLMR:4, even though median aggregation decreased from 138 to 91 (Carrascosa-Zamacois et al., 2022).
This produces an important correction to the common assumption that “more concurrent links always imply lower delay.” Under high-load symmetric contention with too few orthogonal channels, unrestricted multi-radio concurrency can create starvation and worse tail latency than a well-planned single-link channelization. Several papers therefore argue that the best configuration in dense environments may involve limiting concurrency, assigning only one link per flow, or guaranteeing at least one contention-free channel per BSS (Carrascosa-Zamacois et al., 2022, Carrascosa-Zamacois et al., 2022, López-Raventós et al., 2021).
Another anomaly appears under asymmetric link occupancies. Experimental studies driven by 5 GHz occupancy traces showed that when both links were symmetrically occupied, MLO-STR could reduce latency by one order of magnitude, whereas in asymmetrically occupied channels it could be detrimental and even increase latency. The mechanism identified was premature packet mapping: if a packet is committed to the busier link before backoff completion, repeated backoff interruption on that link can dominate the access delay. Opportunistic STR or STR+, which defers link assignment until a backoff actually expires, removed this pathology and guaranteed latency no worse than the better single link (Carrascosa et al., 2021, Carrascosa et al., 2023).
Coexistence with legacy Wi-Fi also depends on the chosen MLO flavor. Simulation studies comparing MLMR and MLSR found that MLMR’s potential is most visible in isolated or lightly loaded deployments, whereas MLSR can be more effective at medium to high load when fairness with legacy single-link nodes matters. In flow-level coexistence studies, dynamic MLO allocation improved not only the MLO BSS itself but also legacy multi-band single-link neighbors: the legacy MB-SL BSS median satisfaction increased by approximately when surrounded by MLO BSSs, and the dynamic MCAB policy reduced the negative impact of legacy neighbors on the central MLO BSS by almost compared with MCAA (Carrascosa-Zamacois et al., 2022, López-Raventós et al., 2022).
5. Traffic steering, control, and optimization
Because per-link contention is independent while QoS objectives are end-to-end, MLO requires an explicit traffic manager. Early policy-based work compared Single Link Less Congested Interface (SLCI), Multi Link Congestion-aware Load balancing at flow arrivals (MCAA), and the dynamic Multi Link Congestion-aware Load balancing (MCAB) policy. MCAB reallocates flows based on per-link free airtime and, for long-lasting video flows, improved average satisfaction by versus MCAA and versus SLCI at the $95$0 worst-case tail; it also achieved $95$1 in more than $95$2 of scenarios (López-Raventós et al., 2022).
Learning-based methods generalize this control problem. A federated reinforcement-learning framework for decentralized MLO link activation used a max–min reward shaped by neighboring APs’ worst average reward, $95$3, and improved the average minimum achieved data rate from 156.204 Mbps with independent RL to 193.212 Mbps, outperforming random and fixed schemes as well. The explicit conclusion was that activating fewer links can be preferable in dense overlapping BSSs when it reduces interference and contention (Ali et al., 2023).
Cross-layer formulations make the coupling sharper by optimizing both U-MAC traffic allocation and L-MAC contention parameters. One Wi-Fi 7 study extended the Bianchi model to STR MLO, with per-link attempt probability
$95$4
and used an LSTM-SAC algorithm to jointly optimize traffic-allocation vectors $95$5 and initial contention windows $95$6. In an event-based simulator, that method reduced access delay by $95$7 versus SAC, $95$8 versus LSTM-DDPG, and $95$9 versus an allocation-only variant at 0, while also providing the highest throughput and near-best fairness (Liu et al., 19 Mar 2026).
Optimization also extends to channel assignment and AP coordination. A centralized Wi-Fi 7 framework formulated AP–STA association as a totally unimodular LP and then applied proportional-fair link allocation, reporting gains up to 1 in the abstract and, for some configurations, 2 higher throughput than greedy+PF and 3 higher than greedy+RR (Zhang et al., 2023). For dense dynamic channel allocation, BAI-MCTS cast each global MLO configuration as a bandit arm and converged approximately 4 faster than state-of-the-art algorithms at 5 of optimum; the LLM-BAI-MCTS extension increased convergence rate by over 6 in dense networks (Lian et al., 5 Jun 2025).
The same optimization logic appears at the EDCA layer. A GA-based MLO EDCA method jointly assigned access categories to links and tuned AIFS, CW, TXOP, and retry limits to minimize overall loss subject to delay-violation constraints. The paper’s analytical model linked per-link collision probabilities, zone-based AIFS interactions, and delay CCDFs, showing that static EDCA parameters are inadequate when AC-to-link assignments change the contention landscape (Yi et al., 30 Sep 2025).
6. Implementations, extensions, and prospective directions
MLO research increasingly includes experimental platforms rather than only analysis and simulation. A software platform for industrial Wi-Fi built a Virtual MLD from commodity PCs, dual PCIe Wi-Fi adapters, ath9k, and modified mac80211 hooks that exposed per-link TX status to user space. In more than a week of measurements over four non-overlapping 2.4 GHz channels, outcome correlations between most channel pairs were near zero, supporting the view that cross-link diversity is real and exploitable for duplication or steering in industrial settings (Rosani et al., 2024).
At higher frequencies, MLO is also being used to rethink control-plane design. In mm-wave WLANs, cross-link RTS/CTS uses a sub-7 GHz link for omnidirectional control while reserving mm-wave time for directional data. Analytical and numerical results show a clear trade-off: cross-link RTS/CTS increases success probability and reduces hidden terminals, but lowers network throughput relative to purely directional RTS/CTS because it suppresses spatial reuse more aggressively (Chen et al., 7 Nov 2025).
Another proposed extension is Continuous Multi-Link Operation (ConMLO), which uses independent channel-access attempts across all addressable links so that at least one acquired link is ready for transmission at all times. The abstract reports that ConMLO can guarantee continuous channel acquisition under different occupancy scenarios without compromising fairness relative to legacy single-link operation and baseline MLO (Fontanesi et al., 2024).
Several forward-looking papers explicitly connect MLO to Wi-Fi 8 or 802.11bn directions. Proposed pairings include learning-based control for concurrency and channel reuse, operation in 60 GHz where available spectrum could increase by nearly one order of magnitude, coordinated beamforming to suppress inter-BSS interference, and Ultra High Reliability mechanisms. This suggests that MLO is becoming less a single feature than a coordination substrate onto which future unlicensed reliability and multi-band scheduling mechanisms will be attached (Carrascosa-Zamacois et al., 2022).
Open questions remain consistent across the literature: dynamic channel assignment under QoS constraints, real-time adaptation of concurrency, interaction with OFDMA, MU-MIMO, puncturing, and multi-RU scheduling, operation under heterogeneous traffic and heterogeneous device capabilities, and standards-compliant coexistence between STR-capable MLDs, NSTR MLDs, and legacy nodes. The research record therefore treats MLO not as an automatically beneficial replacement for single-link operation, but as a high-dimensional MAC framework whose performance is contingent on careful control of traffic placement, concurrency, and contention.