What is TCP BBR

Google's TCP flow control BBR slows down third-party downloads

Since the Internet has existed, researchers have been working on better and better variants of TCP flow control, because the lines are getting faster and faster and require new adjustments. A new concept comes from Google with Bottleneck, Bandwidth and Round Trip Time (BBR). However, as tests at the recent RIPE meeting in Marseille show, BBR increases throughput at the expense of other Internet users.

Google has been using BBR since last summer. According to studies by the Karlsruhe Institute of Technology (KIT), the implementation has been running on all servers at Google.com and YouTube.

Alternating braking and accelerating helps

Generally, a sender, such as a web server, knows little about the route to the receiver - it knows neither the number of routers nor their maximum data rate, and certainly not their current load. However, a TCP control does not need to know these details: it is sufficient to determine the latency and accelerate it during transmission until a packet in the stream is lost. Then the transmitter throttles and gradually accelerates again. Throttling and careful acceleration are particularly important for ensuring that the capacity of a line is distributed at least fairly fairly among all current users. Because once a router is overloaded, its input buffers are full and it helps immediately if all servers that are currently sending data via this router temporarily send more slowly. It then empties its full buffers so that fewer packets are lost as a result. This then benefits all participants, the resource is shared fraternally.

However, BBR will continue to fire at high speed for a while if packets are lost. It uses fast lines better than conventional TCP flow controls. Geoff Huston, chief scientist of the Asia-Pacific Network Information Center (APNIC) has been analyzing BBR for a while and presented the results of his measurements at the RIPE meeting. According to this, BBR flattens streams running in parallel on the same route by usurping the bandwidth completely. Huston compared BBR with the TCP flow control "Cubic", which is widespread on Linux, and spoke of a "brutal clash" of unequal competitors. BBR completely booted out Cubic in the test and pulled the bandwidth into itself. BBR gave him a stable 400 Mbit / s on a route from Australia to Germany, reports Huston.

BBR cannibalizes itself too

For the Australian, Google's flow control is a "perhaps necessary answer on the way to terabit speeds on the Internet", but it calls into question the traffic management of network operators. So far, BBR has only worked well for individual servers because it is still under development and is therefore still very rarely used. Measurements show, however, that if several servers have to compete for bandwidth with BBR, they hinder each other more than servers that use conventional, cooperative TCP flow control.

KIT took a closer look at what happens when BBR competes against BBR. Because BBR lacks an explicit fairness mechanism, races between BBR users are also unfair. "We measured situations in which one BBR flow used 10 percent, while another used 90 percent of the bandwidth," says Roland Bless from the Institute for Telematics at KIT.

Overload situations inevitably

Bless therefore calls the use of BBR technology "premature". "The effects of the BBR process have not yet been adequately investigated, in particular not by an independent party," explained Bless. However, the KIT also warns of a face-off effect. It is true that users of Google services could benefit from BBR in some situations. At the same time, however, others would be too severely disadvantaged. The tricky thing from Bless' point of view is that Google simply applies the process to each of its broadcasters. "In the case of several transmitters, this inevitably leads to an overload situation."

In his opinion, the fact that no one has complained on the Internet so far is due to the fact that the applications - the YouTube stream or the Google search - limit the transmission rate by themselves. "The negative effects on other data streams are presumably not observable at the moment," explained Bless.

Huston is now particularly excited to see how BBR will develop. Google is working on a new version. Bless, however, is skeptical of the announcement for BBR2. Google has promised to rein in the BBR2. But the Karlsruhe expert doubts that it would be a good idea to "tinker with a mechanism whose model assumptions are fundamentally wrong". (dz)

Read comments (136) Go to the homepage
Ad ad