How throughput enhancements dramatically boost 802.11n MAC efficiency--Part II

The primary method used to improve the MAC performance is to amortize the high cost of medium access over a larger number of data frames. First, 802.11n incorporates the mechanisms introduced in 802.11e, a prior amendment to the standard. Though these mechanisms were devised to provide differentiated QoS to MAC users, they also help amortize some of the MAC overheads. It introduced the concept of a Transmit Opportunity (TxOP), whereby a station that acquires the medium, does so for a bounded time period (as opposed to a single frame-ack sequence in the original DCF.) Thus the DIFS wait and backoff countdown steps are required only once in every TxOP duration. Another scheme introduced is the Block Acknowledgement (BA.) Instead of each frame being individually acknowledged, a set of frames may be acknowledged using a BA response. This amortizes the response overhead, over a larger number of data frames. These improvements are shown in the first two rows of Figure (3).

802.11n introduced a further set of improvements, illustrated in the last 3 rows of Figure (3.) The first improvement is to reduce the interframe space between succesive transmissions. SIFS was designed to allow for a station to turn from receive mode to transmit mode. Since the station remains in transmit mode for the duration of TxOP, when using the BA scheme, the interframe gap between these transmissions may be reduced. The reduced interframe space (RIFS) is designed to allow a receiver to get ready for frame reception. The next improvement is to concatenate the data frames in the burst, so as to do away with the interframe space and PHY headers. This process is called Frame Aggregation, and is the major throughput enhancing mechanism introduced in 802.11n. Finally, a separate BA Request was deemed redundant for the purpose of soliciting a BA. In 802.11n, the BA request is implicit for aggregated frame.

The next three sections discuss these MAC enhancements in detail. Section (5) will describe the frame aggregation schemes. Section (6) explains the block acknowledgment mechanisms. And in section (7) we discuss an enhancement to the channel access rules to allow for bi-directional TxOPs.

Aggregation

802.11n employs two steps of accumulation to increase the size of the data frame. The first, which is at the top of the MAC, collates MAC service data units (MSDUs) and is called A-MSDU. Another, at the bottom of the MAC, accumulates MAC protocol data units (MPDUs) and is called A-MPDU. Figure (4) shows the location of these aggregation steps relative to the rest of the MAC activities.

A-MSDU

MAC service data units (MSDU) belonging to the same service category, and destined for the same receiver are accumulated to form an aggregate MAC Service data unit (A-MSDU.) Each MSDU is prepended with a subframe header consisting of the destination address, source address and payload length. The data unit is also padded to a 32-bit word boundary. These units are then concatenated and passed on to the rest of the MAC protocol stack as shown in Figure (5). The aggregation process is limited by the maximum length of A-MSDU that the receiver can process. This can be 3839 bytes or 7935 bytes, and is known from the HT capabilities information declared by each station. The aggregation process is also limited by the requirements of MSDU lifetime. Each MSDU has a designated lifetime, and the MAC must attempt to deliver the MSDU to the peer-LLC within this time, with high probability. The process of aggregation involves waiting for more MSDUs to come from the LLC. The aggregation process must be limited such that the rest of the MAC steps and transmission, can meet the delivery timing requirement.

A-MPDU

MAC protocol data units (MPDUs) at the bottom of the MAC are accumulated into a single aggregate protocol data unit (A-MPDU.) The encapsulation process is depicted in figure (6.) A delimiter is prepended to the MPDU, and pad bytes are added to align the end to a 32-bit word boundary. The delimiter contains the MPDU length, 11-n signature (which is the ASCII character 'n'), and a CRC byte. The purpose of the delimiter is to allow the receiver to delineate later MPDUs in the aggregate even if one of the preceding delimiers is corrupted. All MPDUs in the aggregate must belong to the same service category and be addressed to the same receiver. The total length of the aggregate is limited to 8K, 16K, 32K or 64K depending on the capability of the receiver. Since the aggregate is transmitted in one PHY frame, the receiver is trained only once at the beginning of the frame. The channel conditions may change significantly during a long frame, causing errors in reception. Hence the length must also be limited by the channel coherence duration. Further, a maximum of 64 MPDUs may be aggregated due to the limitation of the block aknowledgement response frame. Since MPDU encapsulation, and A-MPDU aggregation happen in the network interface, the amount of aggregation is also limited by the number of frames  available in the transmitter queue.

Two level aggregation

Two level aggregation involves both A-MSDU and A-MPDU accumulation processes. This is illustrated in figure (7). Support for this feature is optional, and must be pre-negotiated on a link. This type of aggregation helps in the anomalous case where a succession of small MSDUs must be transported. If only A-MPDU aggregation is used, the limit of 64 MPDUs may result in a portion of the transmit opportunity to remain unused. A similar situation may arise with larger packets at higher data rates.

The two types of aggregation schemes in 802.11n differ in their utility. A-MSDU aggregation may be performed in the host system, which allows it to wait for MSDUs to arrive. By contrast the ability of A-MPDU aggregation, to wait for MPDUs, is limited by the space available in the transmit queue of the network interface. An A-MSDU is encapsulated into a single MPDU, with one sequence number. Should the MPDU fail to be received, it must be retransmitted fully. On the other hand, MPDUs in an A-MPDU have separate delimiters, allowing for partial recovery at the receiver, and unique sequence numbers allowing for selective retransmission. Finally, A-MSDU aggregation is more efficient than A-MPDU aggregation, as the latter must append a MPDU header to each subframe.

Block Acknowledgment Mechanism

The Block Acknowledge mechanism, initially introduced in 802.11e amendment, has been changed and improved in 802.11n.

Implicit Block Acknowledgment

The first improvement is to remove the requirement of a BA request (BAR) to solicit a BA response. Instead it was deemed that the BA is the normal response to an A-MPDU. This sequence is shown in figure (8) part (a). This step removes the overhead of the BAR, and also eliminates the single point of failure when the BA request is not properly received. The 802.11n BA mechanism also allows multiple A-MPDUs to be acknowledged using a single BA. In this case, all except the last A-MPDU are sent with "block acknowledge" policy directing the receiver to only record the reception status of their MPDUs. The last A-MPDU is sent with "normal acknowledge" policy, which implicitly solicits the BA for all MPDUs sent since the last BA. Alternately all A-MPDUs could be sent with "block acknowledge" policy, and a separate BAR may explicitly request for the BA. These sequences are illustrated in figure (8), parts (b) and (c).

Compressed Block Acknowledgment

It was recognized that the mechanisms of frame fragmentation and frame aggregation are not likely to be useful at the same time. The purpose of fragmentation is to break MSDUs into smaller pieces for transmission over a noisy channel. This process improves the chance that the fragments will be received correctly, and hence reduces the retransmission overhead. Aggregation is used when the channel is good enough to permit transfers at high data rates, and increases the payload size to take advantage of that. With this understanding, 802.11n does not allow fragmented MSDUs to be part of an A-MPDU. The BA response now requires only 1 bit per MSDU. Thus the compressed BA payload is only 64-bits long, compared to the 1024 bits for the full BA. It occupies less memory at the receiver,  takes lesser time on air, and may be sent using more robust modulation

Partial State Block Acknowledgment

The 802.11e block acknowledgment system requires the recipient to maintain a MPDU reception scoreboard for all block acknowledge sessions in progress. Since the BA response must be produced on demand and transmitted immediately, this presents an implementation challenge in terms of memory requirement, and speed of access of such data structures. 802.11n relieves the recipient of this requirement, by allowing it to store the reception scorecard of a few most recent sessions. It is the burden of the transmitter to retrieve the BA response prior to it being discarded (and overwritten) at the receiver. If the scoreboard entry at the receiver has been discarded at the time a request is received, it must return a BA response with all bits set to zero. This will cause a retransmission of all MPDUs. The typical usage scenario is where the transmitter sends one or more A-MPDUs to a receiver, and retrieves the BA response within the same TxOP as shown in Figure (8).

Reverse Direction Protocol

Several applications, such as FTP and HTTP, result in highly asymmetrical traffic, that is the traffic in the forward direction is much more than the traffic in the reverse. However the application transfer speed is strongly dependent on the latency of the transfers in the reverse direction as well. Under the normal access rules, the recipient must obtain a TxOP and transfer the response frames. This results in two issues - first, there is a delay in transmission of the responses, and secondly the TxOP in the reverse direction is not used fully. Both these problems have been alleviated in 802.11n by the introduction of the Reverse Direction Protocol (RDP.) This allows for a "bi-directional TxOP", in the sense that the holder of a TxOP may "sub-lease" a part of its TxOP to another station. This can be used in two ways - the transmitter in the forward direction may make room for its peer to send a response frame at the tail end of its TxOP, or the recipient may obtain its own TxOP and hand it over to its peer, after transmitting the response frames. This simple enhancement to the rules of medium access require very little signaling, but experimental simulations have shown an improvement of upto 40% for TCP throughput.   

Control and Use of PHY features

Finally, it must be mentioned that the MAC layer must carry out several processes to enable and optimally use the various PHY layer enhancements.

The first and foremost of these roles is in obtaining the steering matrix to be used for transmit beamforming. When explicit beamforming is used, the transmitter must initiate a handshake, from which the recipient must compute and respond with the channel state information as soon as possible. When employing implicit beamforming, the station must request its peer to send a transmission that will enable it to estimate the channel. In addition it must also correct for analog discrepancies between its transmit and receive paths. The transmitter must store this channel information, and use it to direct transmissions to this recipient.

802.11n provides 77 different schemes (MCS) for transmission, including variations in number of streams, modulation scheme, coding rate, guard interval, and channel bandwidth. Of these various permutations, a subset will be allowed for the link - based on what is supported by  both transmitter and the receiver. The MAC must attempt to seek the best option for a given frame, and communicate it to the PHY. The choice is not static, and must be adapted over time. Prior to 802.11n, rate selection was an open loop problem, with the transmitter deciding the best option. 802.11n MAC provides a MCS feedback protocol, to allow the transmitter to query the receiver before deciding on the optimal MCS.

Also, the MAC must also make a proper selection of the level and type of aggregation to employ during transmission. This primarily involves a delay-efficiency trade-off, but the length of the frame must also be suited well to the channel conditions.

And finally, the 802.11n MAC also supports interoperability with legacy devices, and co-existence with devices which only support 20 MHz channel widths. The protocol has added features to optimize the throughput in the presence of legacy devices, and devices with different levels of 802.11n capabilities.

The activities mentioned in this section are  critical to allow the full potential of the PHY improvements to be realized by applications utilizing the wireless link.

Conclusion

With this we conclude our presentation of the  MAC layer enhancements in 802.11n. Like the previous amendments, 802.11a and 802.11g, the PHY data rates in this amendment have been improved by an order of magnitude. Several streams are simultaneously transmitted over multiple antennas, using a wider spectrum to achieve this phenomenal gain. In order to translate this gain into usable bandwidth for applications, the MAC protocol was amended to reduce or amortize the access and transmission overheads. The primary mechanisms of aggregation, block acknowledgement and bi-direction transmit opportunity were introduced. Besides, the MAC provides mechanisms for the proper use and control of the new PHY features. With these enhancements to MAC and PHY, 802.11n achieves a maximum data rate of 600 Mbps, and a maximum throughput of around 375 Mbps above the MAC. Now the stage is set for the 802.11 to replace wired ethernet as the dominant LAN access technology.

About the Author

Probir Sarkar has worked on 802.11 MAC architecture, design and verification at Redpine Signals in Hyderabad, India. Redpine Signals is a wireless start-up developing silicon solutions for advanced wireless protocols, and system solutions for selected wireless applications. He is currently working on Processor Architecture at ARM, Bangalore.