Guest Blogger: AirTight Networks
Author: Bhaskaran Raman, PhD.
Series Editor: Tom Carpenter
About the WiSE article series: Wireless is inherently complex; its study spans at least two engineering disciplines: Electrical Engineering and Computer Science. Add to this the nuances of various standards, vendor implementations, RF environments, and protocol interactions, and it is not uncommon to feel a little lost in understanding the various aspects of Wi-Fi network operation. In this series of short articles, we explain various Wi-Fi subtleties, to work toward a better understanding of Wi-Fi network deployments.
WiSE Article No.2
Wi-Fi Throughput Measurement – Parameters that Matter
The first WiSE article talked about the ballpark range of numbers to expect, while measuring Wi-Fi throughput. The measurement process involves a wide range of settings: in the AP, the client, as well as the measurement software and the environment. Suppose you end up with a number less than expected, what parameter settings should you check? Beyond the obvious issues such as interference and signal strength, this article explains other important parameter settings that significantly affect the measured throughput. It also provides some common parameters that are less important, which you don’t need to tweak in most cases. The provided figure gives a preview of the various parameters, and the potential throughput reduction under non-optimal parameter settings; the details are discussed below.
Parameter settings with large effect on throughput
Two factors are well known for their impact on the measured throughput: (1) Interference from neighboring Wi-Fi and non-Wi-Fi devices, and (2) Signal strength between the AP and the client under test; or roughly equivalently, the distance between them. However, there are additional factors which we now describe.
Packet aggregation: Many APs and clients have a configuration parameter to enable or disable packet aggregation. This usually refers to physical layer aggregation (A-MPDU) in the HT PHY. Enabling this aggregation is extremely important to realize the benefits of the high data rates offered by 802.11n (HT). We can use the algebra from the first WiSE article to estimate that without packet aggregation, throughput can go down by a factor of 5 or more!
TCP parameter settings: TCP settings are often non-obvious, yet are important parameters. Suppose you run a throughput test using measurement software such as iperf. The throughput numbers you get using default settings can surprise you! Here it is imperative to explain an important operational mechanism in TCP: the window-based acknowledgment (ACK) scheme. The TCP data sender maintains a window of packets in its buffer. The mechanism mandates that the sender send only one window worth of packets, and not more, before getting an ACK from the receiver.
Thus the window size, otherwise called the operating system’s socket buffer size, can determine how fast the TCP sender can send data. As it turns out, the default window size is 64KB or 128KB in most operating systems. Also, depending on the compilation and configuration of the measurement software, the default can even be as low as 8KB. In such a scenario, you can end up with a throughput number which is just 15-20% of what the throughput algebra tells you!
To address this, there are two options while measuring throughput on the wireless link: (a) increase the TCP window size, or (b) use multiple parallel TCP streams. It is generally advisable to use both these options simultaneously. The figure below shows the downlink TCP throughput numbers, measured in the AirTight Networks Wi-Fi lab for different windows sizes, and different numbers of parallel TCP streams; uplink measurements showed similar trends. Individual numbers may vary in other settings, but the take-away here is that the TCP window size and the number of parallel streams are important parameters that need attention.
Number of MIMO streams: Throughput with the 2×2 data rates will clearly be lower than with the 3×3 data rates. Note when we say NxN here, it means both AP and client are NxN. It is however interesting to note that 2×2 operation can give as much as 80%, not just 67% as the physical layer data rates may imply, of the throughput with 3×3 operation. The reason for these is that the fixed overheads have a larger effect at the higher data rates possible with higher number of MIMO streams. This is because, as you would recall from first WISE article, much of the fixed overhead comes from factors unrelated to the data rates, such as the backoff gap.
Channel bonding and 5 GHz operation: It is also clear that using a 20 MHz channel width will give much less throughput than 40 MHz operation. Interestingly again, the TCP throughput with 20 MHz operation is 67% of the 40 MHz throughput, not just 50% as the physical data rate may imply. This again is due to the fixed overheads unrelated to the data rates.
It is also worth noting that while operating in the 2.4 GHz band, channel bonding to achieve 40 MHz operation is almost never possible, because you will always find some neighboring AP which overlaps with your channel which makes standard compliant devices to fall back to 20 MHz. This means that the ‘802.11an’ setting is necessary in practically all scenarios to achieve 40 MHz operation.
Parameter settings with relatively little effect on throughput
While analyzing throughput measurements, it is also important to know what parameters not to concern yourself about. We list below, three common parameter settings which have only a small effect (less than 10%) on the measured throughput.
Guard Interval (GI): Many APs and clients allow configuration of the use of short guard interval. The guard interval is that between successive symbols in transmission; the long value for this is 800 ns and the short value is 400 ns. Typically, the use of long GI results in 10% lesser physical data rate, which in turn translates to just 5-6% reduction in TCP throughput compared to the use of short GI.
Long preamble: Not all clients/APs in a deployment may be 802.11n capable. For co-existence among pre-802.11n radios, 802.11n defines the use of a long preamble in the physical layer. This is also called HT-mixed preamble, while the short version of the preamble is called HT-greenfield. Although this is given as a setting in many commercial APs and clients, this setting matters little from a throughput perspective. The long preamble adds only an extra 12 micro-seconds to the packet length, which in turn translates to less that 3% reduction in TCP throughput.
Encryption mode: Different encryption modes add different amounts of header overheads: WEP adds 8 bytes, CCMP adds 16 bytes, while TKIP adds 20 bytes. These are negligible for TCP throughput since the overall packet size is close to the Ethernet MTU of 1500 bytes. Thus the throughput reduction is under 2%. It is also worth noting that encryption and decryption at the link layer do not slow down the link, since most implementations today encrypt/decrypt in hardware at the link speed.
- Like in the first article, we have talked only about single client throughput measurement
- While this article lists the important parameter settings, this list is not exhaustive
Bhaskaran Raman is a scientist at AirTight Networks working on high performance Wi-Fi architecture. Bhaskar received his B.Tech in Computer Science & Engineering from IIT Madras in May 1997 and his M.S. and Ph.D. in Computer Science from University of California, Berkeley, in 1999 and 2002 respectively. He was on the faculty in the CSE department at IIT Kanpur (India) from 2003-07. Since July 2007, he has been a professor at the CSE department at IIT Bombay. His research interests and expertise are in communication networks, more specifically in wireless & mobile networks.