Multi-Link Operation (MLO) Modes
802.11be defines four ways a Multi-Link Device (MLD) can use two or more links at once. Pick a mode below and watch the same traffic scenario play out on 5 GHz and 6 GHz. The differences between MLSR, EMLSR, STR, and NSTR become visible the moment the links have to share time.
— Shankar K. · Source: IEEE 802.11be-2024 Section 35.3, 9.4.2.312 · Wireshark: wlan.eht.multi_link
A Wi-Fi 7 device with MLO can use 2 bands (say 5 GHz + 6 GHz) at the same time against the same AP. That device is called a Multi-Link Device. How it uses those bands — one at a time, in parallel, or with constraints — is the mode. The four modes below trade throughput for cost, complexity, and power. Scroll to the bottom for the Basic / Intermediate / Expert guide if the acronyms are new.
wlan.eht.multi_link expands the tree; the EML Capabilities subfield is at wlan.eht.multi_link.common_info.eml_capabilities.
The modes above show how an MLD uses two links. Whether that delivers lower latency or higher throughput is a separate question, and the research answer is: it depends, on four things. MLO is a conditional optimization, not a free win. Get the conditions wrong and a second link can leave latency unchanged, or make it worse.
- Device class. STR delivers the real latency win. EMLSR delivers resilience but peak throughput bounded by one radio. A device that advertises MLO without simultaneous-link capability barely moves 95th-percentile latency at all.
- Channel-occupancy symmetry. When both links are similarly occupied, MLO shines. When one link is much busier than the other, STR can backfire.
- Load and contention. In an isolated or lightly-loaded BSS, MLO reduces delay. In a dense, contended deployment, multi-radio MLO can starve neighbors and increase worst-case delay — the MLO delay anomaly.
- Number of links. Two links capture most of the benefit. A third and fourth give rapidly diminishing returns, and under contention can hurt.
An enterprise upgrades a dense office floor to Wi-Fi 7 APs and enables MLO across 5 GHz and 6 GHz for everything, expecting lower latency for voice and real-time apps. After rollout, 95th-percentile latency on the busiest part of the floor is higher than it was on the Wi-Fi 6 single-link baseline. Average latency looks fine. The worst-case got worse.
This is the MLO delay anomaly (Carrascosa-Zamacois et al., arXiv:2210.07695). Under STR multi-radio operation in a contended environment, each MLD aggressively grabs whatever links are free. When several MLDs do this at once, they block one another: a device can find all of its links busy a meaningful fraction of the time and defer its backoff on every one of them. The paper's enterprise model shows that in this regime, aggressive multi-radio access (STR EMLMR) with more links can produce higher worst-case delay than a well-planned single-link assignment. Greedy multi-link access lowers best-case delay but can raise worst-case delay.
The fix the research lands on is structural: guarantee each BSS at least one contention-free link. A "1+1" arrangement, one reserved channel plus one shared, beats even an overprovisioned five-radio setup, because the reserved link removes the starvation case entirely. In a capture, the signal to look for is MLDs operating multiple links under high channel occupancy with rising 95th-percentile delay, the opposite of the marketing promise. The mode and capability negotiation in the Multi-Link Element tells you which devices are STR-capable and how many simultaneous links they intend to use (see the expert tier below for the exact fields).
Most of these results come from analytical models and simulation; the latency study (arXiv:2111.09281) is the one grounded in real over-the-air 5 GHz occupancy traces. Treat the magnitudes as direction-of-effect, not field-calibrated guarantees, and confirm against your own captures. The anomaly is a contention phenomenon: it does not appear in an isolated BSS, where MLO behaves exactly as the brochure says.
A second link helps most when the air is busy and both bands are similarly occupied. If one band is far busier than the other, or if your device is EMLSR rather than STR, the second link does much less than the headline number suggests. The real, reliable benefit of MLO for most client devices is not peak speed, it is resilience: when one band degrades, the device keeps working on the other without re-connecting.
Three findings from the UPF Barcelona MLO studies:
Channel symmetry decides the latency win. With two symmetrically-occupied links, STR improved 95th-percentile latency by up to roughly an order of magnitude (arXiv:2111.09281). But with asymmetric occupancy, where the chosen secondary link is much busier, STR can instead increase worst-case latency versus single-link, because it commits a packet to a channel that then goes busy and freezes its backoff. The proposed fix, opportunistic STR (STR+), defers committing a packet until a backoff actually expires, guaranteeing delay no worse than single-link.
Single-radio MLO underdelivers on worst-case latency. A device that can only use its secondary link when the primary is also idle provides negligible 95th-percentile improvement over single-link at medium-to-high load. If a device advertises MLO but cannot transmit on links simultaneously, expect single-link-class worst-case latency regardless of the marketing. As of 2026, shipping Wi-Fi 7 clients are typically either STR-capable MLMR or single-radio EMLSR.
Diminishing returns past two links. The finite-load delay model (arXiv:2212.12420) shows that most of the worst-case-delay benefit comes from the first additional link; a third and fourth add little, because the probability that a packet even needs buffering vanishes after one or two extra links. Throughput at a fixed delay target does keep scaling with more links under high contention, but worst-case delay does not keep improving.
The starvation anomaly (arXiv:2210.07695). In contention-free conditions, STR EMLMR reduces tail delay monotonically as links are added. Under high load and contention the result inverts: multi-link devices block contending BSSs, and adding links can raise delay rather than lower it. The durable fix is channel assignment, not more radios: a 1+1 design (one contention-free reserved channel plus one shared) outperformed a five-radio configuration at high load, because it guarantees an escape link per BSS. EMLSR, by using one link at a time, keeps delay at or below single-link across all loads.
Coexistence with legacy single-link (arXiv:2205.15065). A multi-radio MLO BSS boosts its own throughput largely at the expense of neighboring single-link BSSs, taking airtime from them. When a single-link neighbor is heavily loaded, aggressive multi-radio MLO can starve it. In medium-to-high-load mixed environments, the simpler single-radio mode is both fairer to neighbors and competitive on the multi-link device's own latency — an argument against assuming the most aggressive mode is the best one.
Where the throughput story still holds. The conditional cautions above are about worst-case delay under contention. For aggregate throughput, multi-link still wins clearly, and the experimental latency-and-throughput study (arXiv:2305.02052) measures both sides of that trade. The takeaway is not "MLO is overrated," it is "MLO is a delay optimization whose sign depends on contention and symmetry, while its throughput benefit is robust."
What to verify in a capture. Mode and capability negotiation live in the Association Request's Multi-Link Element. Base filter to expand the element: wlan.eht.multi_link. Inside it:
- Common Info → EML Capabilities carries the EMLSR Support and EMLMR Support bits (whether the device can do enhanced single-radio or enhanced multi-radio operation).
- MLD Capabilities → Maximum Number of Simultaneous Links is how you read STR. A non-zero value means the device is MLMR (multi-radio, STR-capable); zero means single-radio. The field is offset by one, so the actual number of simultaneous links is value + 1.
- MLSR is mandatory for every AP and non-AP MLD, so its presence tells you nothing; EMLSR and STR support are the differentiators that decide which latency curve above applies.
A device advertising MLO with "Maximum Number of Simultaneous Links = 0" is single-radio (EMLSR or MLSR), and its worst-case-latency math is closer to Wi-Fi 6E than the headline.
MLO is not a speed button. It is a resilience feature for most clients and a delay optimization for STR-capable ones, and under dense contention an aggressive multi-radio configuration can raise worst-case latency rather than lower it. Read the Multi-Link Element before you trust the brochure: the capability fields tell you which of the curves above your device is actually on.
A new Wi-Fi 7 flagship phone on a Wi-Fi 7 mesh router. Speed test next to the router: 2 Gbps. Walk into the next room: 800 Mbps. Living room: 400 Mbps. Reviewers called this "game-changing." The user is unimpressed.
The 2 Gbps headline assumes the device is running MLO in Simultaneous Transmit/Receive (STR) mode — two radios powered at the same time, on 5 GHz and 6 GHz, aggregating traffic. STR requires both radios at close to maximum power continuously, which is battery-hostile. Most client devices, including most flagship phones, actually run enhanced Multi-Link Single Radio (eMLSR), which rapidly switches between bands instead of running them simultaneously. eMLSR keeps the authentication context across bands so handoff is seamless, but peak throughput is bounded by one radio at a time.
Even in eMLSR mode, MLO delivers a real user-visible benefit: when the 5 GHz link degrades mid-session, the device can switch to 6 GHz in milliseconds without re-authenticating or renegotiating a handshake. The user experiences "things just keep working" instead of "Zoom call paused for 3 seconds." The real-world win isn't peak throughput, it's resilience.
Headline Wi-Fi 7 throughput numbers in reviews often assume STR. Before buying a device specifically for MLO aggregate throughput, check the datasheet for "STR capable" or equivalent language. If it says MLO but not STR, it's eMLSR, and the peak-throughput math is closer to Wi-Fi 6E than the marketing suggests.
Multi-Link Operation (MLO) lets a Wi-Fi 7 device use more than one radio channel at the same time against the same access point. Before Wi-Fi 7, a laptop on 5 GHz couldn't also use 6 GHz for the same session — the two bands operated independently, with separate associations and no coordination.
A link is one band + channel combination. For example, "5 GHz Channel 36 at 80 MHz" is one link. A device that supports MLO has two or more of these links active at once.
A Multi-Link Device (MLD) is the logical device that owns those links. To applications and the network layer it looks like one device with one MAC address, even though underneath it may have two or three physical radios. That single identity is what makes MLO clean: no roaming between bands, no double association, no separate IP per band.
How the AP knows which mode a client supports: the STA's Association Request carries a Multi-Link Element. Inside, the EML Capabilities subfield advertises EMLSR and/or EMLMR support. The MLD Capabilities subfield carries the STR capability bit and the maximum simultaneous links. The AP inspects these on day one and provisions the exchange accordingly.
TID-to-link mapping: after association, the AP and STA negotiate which TIDs are served by which links via the Multi-Link Element TID-to-Link Mapping subelement. Voice (TID 6–7) typically maps to the low-latency link; best-effort (TID 0–4) to the high-throughput link. The UMAC at the sender distributes each outgoing frame to the correct LMAC before any EDCA backoff. This is why MLO latency reduction is structural, not probabilistic.
EML Padding Delay. Minimum MAC padding duration the AP adds to the Initial Control Frame (ICF) directed at an EMLSR STA. Gives the STA's PHY time to collapse NSS from the listen configuration (1 SS per link) to the full transmit configuration on the target link. Negotiated during association via the EML Capabilities subfield.
EML Transition Delay. The interval between ICF reception and the first PPDU exchange. Covers the STA's radio retuning and chain reconfiguration. Upper-bounded by the EML Transition Timeout.
Medium Sync Delay Information. An NSTR timer that bounds how long an MLD may hold a TXOP on one link while the other link is in a synchronization-pending state. Prevents one link from starving the other. Present in the Common Info field of the Basic MLE when relevant.
IDC power leakage. In-device coexistence between co-located 5 GHz and 6 GHz radios typically achieves 40-60 dB of isolation. STR requires closer to 80 dB of isolation to transmit one link while receiving another without desensitization. NSTR accepts the lower isolation budget and mandates same-direction operation to avoid the cross-direction saturation case.
UMAC/LMAC split. IEEE 802.11be Section 35.3 formally defines the two-layer MAC for MLO. Each affiliated STA (one per link) consists of a PHY layer and a Lower MAC (LMAC) that handles channel access independently. A single Upper MAC (UMAC) sits above all affiliated STAs, provides the LLC with a single MAC Service Access Point, and owns the MLD MAC (U-MAC) address. The UMAC implements TID-to-link mapping and distributes outgoing frames to the appropriate LMAC before any EDCA backoff occurs on that link.
SR-Tx (Selective Transmit) on UL. In MLO STR mode, the client UMAC selects the transmit link per frame on the uplink — not per session. The selection criterion is the cleanest available band at that instant. ACK frames return on whichever link the AP designates as cleanest, which may be different from the UL path. The TCP window stays open because ACKs continue arriving even when one link is momentarily congested. This is the mechanism behind the WBA Phase 2 UL improvement numbers: +116% UL throughput under CCI is SR-Tx working correctly.
ML-Rx (Multi-Link Receive). The EMLSR client keeps its receiver open on all active links simultaneously at reduced spatial streams (1 SS per link in listen mode). When an Initial Control Frame (ICF / MU-RTS) arrives from the AP on any link, the STA collapses all streams to the target link within the EML Padding Delay window. This is the mechanism that makes EMLSR make-before-break: the receiver never goes dark between links, so there is no gap in connectivity from the TCP layer's perspective.
Per-STA Profile subelement. Length-prefixed inner body that carries per-link association state. "Complete Profile" flag means the inner (Re)AssocReq/Resp body is fully included; "incomplete" means it's compressed against the association link's parameters and only deltas are carried. Link ID in the subelement maps to the 4-bit Link ID in the Common Info.
Wireshark. The MLE tree expands at wlan.eht.multi_link. Per-STA Profile subelements show up as nested trees, each with its own Link ID and inner assoc frame body. On a real capture, the first place to look for mode negotiation evidence is the AssocReq's MLE inside the STA's first AssocReq to the AP MLD.
Building WiFi Analyser V2 · CWNA-109 in progress · one post every two weeks