This article original appeared in the Sept. 2012 issue of Next Gen Mobility.
If you’re the IT director for your household, you’ve probably heard a family member complain, “The Internet is slow today.” This happens with Wi-Fi, 3G/4G and wired services. What’s more, it usually reflects real problems that are present in many parts of the Internet today. The problem is bufferbloat, and it is interesting to see how solutions are being developed and applied in different environments like mobile networks, cable networks and consumer Wi-Fi devices.
Bufferbloat happens when overly large buffers fill up, introducing latency that, in turn, interacts with TCP/IP to dramatically reduce network throughput. Although the term bufferbloat is relatively new, the issue has been known for almost three decades yet never fully resolved. Indeed, every year there’s a new crop of PhD thesis on queuing and buffer management in packet devices.
Ever lower cost memory combined with the mistaken idea that more is better has resulted in the widespread deployment of network devices with excessive buffers. In addition, wireless connections experience dynamically varying bandwidth and delay, making it hard to correctly size buffers in the first place. The result is Internet connections that slow to a crawl intermittently and unnecessarily, and then take minutes to recover.
I first saw this effect on AT&T’s (News - Alert) 3G wireless network in 2009. Everyone was talking about how iPhone data traffic was overloading AT&T’s network but, in many cities, the real problem was excess buffering. During normal operation, latency was less than 200 ms, but as traffic increased, latency suddenly jumped to many seconds, causing all TCP sessions to grind to a halt. It would stay that way for tens of seconds to a minute or more. Jim Getty, who coined the term bufferbloat, has subsequently documented similar problems with residential cable modem services and on links through Wi-Fi routers.
This behavior is unnecessary. With proper buffers, TCP sessions slow down gracefully, providing very acceptable services for far more traffic. Luckily, major players have begun to notice and, although no one acknowledges problems in their specific networks, carriers are obviously working to remove excess buffering as private measurements on several commercial networks show substantial progress.
What’s interesting is how open the process is and how quickly solutions are deployed. For 3G and 4G mobile networks, there is little or no visibility. If the topic has been discussed in 3GPP standards committees, it’s not obvious to an outsider and, as yet, the term bufferbloat doesn’t show up in any 3GPP standard, even though latency has been a major focus for evolving 3G and 4G architectures for over 15 years. In 2009, AT&T Wireless specifically denied its network had any excess buffering, despite clear thirdy party measurements showing repeated instances of 5-8 second ping times. On the other hand, while denying any problems, AT&T appears to have cleaned up its worst buffering issues, based on anecdotal measurements in formerly failing areas. We have no way of knowing what it’s done or how much of the problem the company has solved, but the company appears to have made substantial progress.
Comcast and CableLabs (News - Alert) have been a lot more open. Cable industry PR departments never admitted to any problems, but CableLabs documents describe testing DOCSIS 3 modems and finding severe jitter and very slow webpage downloads. They conclude, “We observed the best web page download performance by using the smallest tested upstream buffer sizes.” Finally, CableLabs has amended some of its specifications to gain configuration control over previously preconfigured buffers. Obviously, it will take time for this equipment to be universally deployed, but already I know of specific Comcast customers in eastern Massachusetts who have seen their worst-case latency to MIT (News - Alert) drop from over 1 second to less than 30 ms.
Consumer Wi-Fi is different. There are just as many problems, but they are spread over many product vendors and many applications. Luckily, specific segments demand solutions. The gaming community is in the forefront. Meanwhile bufferbloat.net has shown how QoS settings can be used to mitigate problems in cable and DSL networks. Thus the Wi-Fi market is fragmented but open and moving to solutions as rapidly as any centrally-controlled network.
Historically, Internet scaling problems are solved at the last possible moment. With bufferbloat, it’s happening again.
Edited by Brooke Neuman