How Google is speeding up the Internet
Google has developed an algorithm that speeds up TCP/IP, the
internet’s primary data transmission protocol, by up to 14%
Engineers at Google have developed a new algorithm to speed
up TCP, the main transport protocol for traffic on the internet, by optimizing
the speed at which traffic is sent so it doesn’t clog up the available routes.
They say their
acceleration method, called bottleneck bandwidth and roundtrip (BBR)
propagation time, measures the fastest way to send data across different routes
and is able to more efficiently handle traffic when data routes become
congested. Google is already using BBR to speed up its YouTube traffic, and
last month the company made BBR available in its Google Cloud Platform. Google
says implementing BBR sped up the already highly-optimized YouTube traffic by
4% on average, and as much as 14% in some countries.
TCP acceleration efforts
TCP developed in the 1970s as part of the protocol suite
TCP/IP to format data into packets for transmission across the internet.
Researchers at the Internet Engineering Task Force (IETF) estimate that more
than 90% of IP traffic is transmitted via TCP.
Over the past few
decades there have been multiple efforts to speed up TCP/IP, many of them
focused on how TCP handles congestion. TCP was designed to slow down how fast
it sends traffic when it senses congestion, which it determines by monitoring
the number of packets lost in transport.
“This worked well for many years because internet switches’
and routers’ small buffers were well-matched to the low-bandwidth of internet
links,” Google explains in a blog post announcing BBR. But so-called “loss-based”
congestion control doesn’t work as well in today’s environments.
Van Jacobson, one of the original authors TCP and one of the
lead engineers who developed BBR, says if TCP only slows down traffic when it
detects packet loss, then it’s too little too late.
“(BBR) is not waiting for a problem to occur, like a loss,”
Jacobson says. “It’s modeling the pipe as if it has a length and diameter to
determine how much data can fit in it.”
Standards
BBR constantly estimates throughput and roundtrip traffic
time across multiple routes, so it knows how long it will take data to traverse
the network if it sends it at a certain rate. By doing so, BBR sends traffic at
a speed the network can handle. This is more effective than the original TCP
congestion controls.
BBR is also compatible with an alternative transport
protocol, quick UDP internet connections (QUIC), devised by Google and being
considered for standardization by the IETF.
BBR is not the first effort to speed up TCP. Researchers at
North Carolina State University are credited with developing one of the most
popular loss-based congestion control algorithms used in TCP today, named
binary increase congestion control (BIC) and subsequently, CUBIC. At a high
level, these also record measurements to estimate the optimal speed at which to
send data when congestion is detected. Another congestion control algorithm
that has become popular is named Reno.
These all use packet loss to determine congestion, though
Jacobson, the Google engineer who developed BBR, says that to his knowledge BBR
is the only TCP algorithm that actually estimates the speed of traffic to
determine the best way to send it, regardless of whether packets have been
lost.
Reaction to BBR
Mirja Kühlewind is a senior researcher at Networked Systems
Group in Zurich and is the IETF’s Transport Area Director who works on TCP
maintenance and improvement. She says creating standards in transport and
congestion control takes a long time. Through dozens of attempts at improving
TCP, there has been only one that has been standardized, and that was before
the development of BIC and BBR.
“In general standardizing congestion control schemes is not
an easy topic,” she says. If any company could push a standard, it could be
Google, given the scale at which they operate, she says.
Jacobson says the company’s goal is for BBR to become a
standard.
Kühlewind says BBR shows promise. “Both Reno and CUBIC work
based on the same principle and react to packet loss as a sign for congestion
and subsequently reduce their sending rate if loss was detected. BBR however
utilizes packet timing information to figure out if the link is
congested.”
Some Google customers are already realizing benefits of BBR.
Wordpress hosts a half-million sites in the Google Cloud and Founder and CTO
Jason Cohen cited Google research which showed BBR provided a 2,700x throughput
improvement compared to other loss-based congestion controls. Queuing delays
were 25x lower, he says.
Users of Google’s Cloud Platform will automatically get the
benefits of BBR when using certain GCP cloud services, including Cloud Spanner,
BigTable, Storage, CDN and Load Balancing at no additional cost.
