Debugging Lower Throughputs

TR-398 is one of the most comprehensive test plans for access point validation, covering scenarios like baseline performance, coverage, multi-client behaviour, stability, latency, roaming etc. While performing TR-398 test suite on a Wi-Fi 6 AP, i started noticing consistent failures in 5 GHz.

For example, consider the 6.2.2 Maximum Throughput Test that intends to measure the maximum throughput performance of the DUT with a single station active. The test uses TCP traffic, and the station is configured for 80MHz in 5GHz. According to the pass/fail criteria defined by Broadband forum, for the configuration of 802.11ax 80MHz 2x2, 880Mbps throughput should be achieved.

To further investigate, i ran throughput tests across all four traffic types: TCP and UDP, in both downlink and uplink directions and, observed quite lower throughputs with TCP traffic. I’ve then repeated the same tests with a reference AP in the same test environment and below are the test results:

Note: Both APs supports only 1G LAN and 2x2 NSS.

As shown above, the 5 GHz TCP downlink throughput is about 100 Mbps lower compared to the reference AP. To identify the cause of this reduced throughput, I captured OTA traffic and analysed the packet capture while running the traffic using LANforge (Candela box).

Packet capture observations:

• More than 90% frames are QoS data frames.

• The RSSI is -35dBm and out of all data frames more than 95% data frames are:
• With higher MCS (MCS 11 in this case – 802.11ax 80Mhz 2x2)
• With shorter guard interval
• With 2x2 NSS
• With 80MHz bandwidth
• No of retry frames = 0.0% (45 frames)

Thereafter I proceeded analysing the data lengths of the QoS data frames

• With the DUT, around 89% of data frames are with data length 3164 bytes and no frames with data length greater than 5000 or 9000 bytes were observed.
• Whereas, the reference AP has around 72% of frames with data length 9260 bytes, indicating significantly larger payload aggregation in the reference AP.

We might wonder how data frame lengths larger than 1500 bytes are observed, and why some frames reach around 3000 and some around 5000 or 9000 bytes.

This is due to frame aggregation in Wi-Fi, which improves efficiency by combining multiple packets into a single transmission. Wi-Fi supports two types of aggregation: A-MSDU and A-MPDU.

1. With A-MSDU [Aggregate MAC Service Data Unit] aggregation, multiple MSDUs are combined into a single A-MSDU. This increases the payload size of a single QoS data frame, resulting in frame lengths larger than 1500 bytes, such as ~3000 bytes, ~4500 bytes, or ~9000 bytes depending on the number of MSDUs aggregated. The aggregated A-MSDU is then transmitted as a single MPDU.
2. With A-MPDU [Aggregate MAC Protocol Data Unit] aggregation, multiple MPDUs are further combined into a single transmission at the PHY layer. This again reduces the overhead by transmitting multiple MPDUs together.

Both the DUT and the reference AP perform A-MSDU and A-MPDU aggregation.

AMSDU Observations:

As shown in the packet capture snapshot below, the DUT uses A-MSDU aggregation with two MSDUs in a single A-MSDU.

While the reference AP aggregates 6 MSDUs per A-MSDU, reducing the overhead and resulting in improved throughput compared to the DUT.

AMPDU Observations:

• A majority of transmissions from both APs are with more than 15 MPDUs in an A-MPDU. However, the reference AP shows ~96% of these frames compared to ~71% for the DUT.
• The DUT shows a noticeable percentage of smaller A-MPDU sizes, with around 12% of frames having fewer than 5 MPDUs, whereas the reference AP shows only ~0.5% in this range.
• This indicates that the reference AP consistently builds larger A-MPDUs, while the DUT transmits comparatively smaller aggregates.
• Larger A-MPDUs reduce contention, and overhead, improving airtime utilization and resulting in higher throughput for the reference AP.

Although both APs support A-MSDU and A-MPDU aggregation, the aggregation behaviour differs noticeably. The DUT mostly aggregates two MSDUs per A-MSDU and shows a quite higher percentage of smaller A-MPDUs, whereas the reference AP aggregates up to six MSDUs per A-MSDU and builds more percentage of larger A-MPDUs.

Due to this, the reference AP transmits more payload in each transmission, reducing per-packet overhead and improving airtime utilization. In contrast, the DUT sends smaller aggregated frames more frequently, which increases overhead and limits throughput.

Even though both APs operate at similar RSSI, MCS, bandwidth, and NSS, the difference in A-MSDU and A-MPDU aggregation efficiency results in the observed throughput with the DUT. [I'm further analysing TCP uplink behaviour, I will share the uplink observations soon].

Have you observed similar scenarios where aggregation significantly impacted throughput, even under similar RSSI and MCS conditions? Let’s discuss.