Godfrey's Team Designs a Parallel Internet with Speed-of-Light Latencies
4/20/2022 8:51:21 AM
At its core, Godfrey and this team believed they could develop an internet with close to speed-of-light latency. Their idea represented not just an improvement to internet performance, but an overhaul to the way it operates and how it is built.
Most know their internet service provider (ISP) provides a certain level of bandwidth, but the speed of a networked application depends on more than that. Even the smallest message takes some time to reach its destination, and for a reply to come back.
“That’s called latency, and it turns out that many interactive applications – like web browsing, cloud gaming, and sometimes even videoconferencing – experience lag that often comes more from latency than bandwidth,” Godfrey said.
Sustaining an energy and belief that internet latency could be greatly improved, Godfrey and the team recently published a paper – entitled “cISP: A Speed-of-Light Internet Service Provider” – that proved their theory possible.
Introducing Godfrey’s Fun Side Project – an Online Escape Room Based on CS Principles
Early in March, Brighten Godfrey also unveiled a side project of his – something he described as “a rather unusual escape room” that incorporates CS principles.
The Missing Links came from Godfrey and co-creator, Renata Cáceres – a professor at the Illinois Wesleyan University School of Music – during peak months of the pandemic, when opportunities for camaraderie and fun were hard to come by. Each enjoyed the typical escape room experience but felt that online versions left a bit to be desired.
“In-person escape rooms offer a whole world of possibilities,” Godfrey said. “You can interact with anything in the room, and there’s a huge amount of freedom and creativity in the puzzle design. One of our ideas was to recreate some of that magic in an online version.”
The story revolves around a team of players going to visit a friend in her lab, until they find out she is missing. While they’re at the lab, an AI agent asks them to assist in fixing the network, at which point they have to discover the links of the network that are missing – hence, the name of the game.
“The puzzles themselves anybody can do, but many of them incorporate computer science concepts,” Godfrey said. “We took inspiration from areas such as graph algorithms, networks, logic and NP-completeness.”
Players have come from around the world, with the fastest “escape” time records currently held by a team from right here at the University of Illinois Urbana-Champaign and a team of engineers from Google.
Godfrey and Cáceres made the collaborative game free to educators with the hope that more students can understand the creativity possible through CS.
“We're hoping that it’s a nice way for people to get together and that it can serve as an activity associated with educational programs in computer science,” Godfrey said. “A lot of software engineers have played with friends, but it’s also great for students, even high schoolers, to play. Once they do, they find a small taste of what CS is about and that it can be fun.”
The Missing Links is not the only science-themed game associated with the University of Illinois Urbana-Champaign. In the in-person escape room LabEscape, players solve puzzles based on principles of physics.
Presented at the 19th USENIX Symposium on Networked Systems Design and Implementation on April 6, the paper showed that “instantiations of cISP across the United States and Europe would achieve mean latencies within 5% of that achievable using great-circle paths at the speed of light, over medium and long distances.”
Godfrey noted this is a very forward-looking project.
“The computing industry develops amazing technology, much of which meets more immediate needs. What I believe academic research shines at – and what I love to pursue – is fundamental architecture innovations. And this project is about as fundamental and architectural as you can get,” Godfrey said. “It’s saying, ‘Let’s build a whole new internet.’”
To design this parallel internet, researchers from Illinois teamed up with collaborators at Yale University, Duke University, and ETH Zürich.
They believe the resulting service could be deployed in 120 of the largest US cities, providing 85% of Americans a much faster internet connection.
Early skepticism stemmed from concerns about feasibility and cost efficiency. The group persevered and answered the questions.
“The important thing that our paper does is that it shows, yes, it is feasible as a service, and that it’s reasonably cost efficient,” Godfrey said. “That allows us to now focus on more exciting things like future applications of a low-latency internet. There may be big opportunities for applications such as distributed augmented reality or extended reality.”
To deliver on such a concept, Godfrey and the team focused on building a new methodology that addresses both bandwidth and latency together.
Bandwidth, the professor said, has improved dramatically on the Internet, but it’s latency that is much more difficult to advance.
“Computing and communication hardware has improved bandwidth, so millions of people can watch 4K UHD video at home whenever they want. But latency depends on physical distance, meaning you can’t improve latency the same way,” Godfrey said. “To come close to the physical limits of the speed of light, and ultimately produce an internet that comes within 5% of the speed of light, was intentionally a very challenging goal.
“It is exciting in retrospect, because you don’t necessarily know where a project like this is going to go when you first start it.”
Where do you start when you have a project that presents an uncertain conclusion?
To begin, the group measured the present internet. One key metric is what’s known as the round-trip time (RTT), which is the time for a device on the internet to send a small packet of information and receive a reply from a remote machine. They found that RTT was typically 3-4 times longer than if messages moved at the speed of light, and often was a factor of 80 or 100 times slower than the speed of light.
Next, the group next identified three primary reasons for this delay.
First, the network doesn’t follow a straight line as the crow flies – instead, it follows the fiber laid underground, which is typically circuitous. Second, ISPs don’t necessarily optimize for latency when they route packets through the network, so packets don’t even follow the shortest path through available fiber. And third, the physics of the material through which information is sent is important: the speed of light in glass – which is what fiber is made of – is about 1.5 times slower than the speed of light in air or in a vacuum.
These factors led to choosing a different physical medium for the speed-of-light internet: wireless directional microwave transmissions between cell towers. This sends information through air instead of glass and makes it easier to construct straight-line paths from tower to tower. But such technology – which has been driven recently by use in high-frequency trading – has much lower bandwidth than fiber.
Their next idea was to use two networks in parallel.
“The present internet does a fine job of delivering bandwidth; it’s the right way to deliver video on demand, for example. Then, a second parallel network is going to be really good at delivering low latency, which is what our new design does,” Godfrey said. “This allows for a system that can send the traffic that’s latency sensitive intelligently over the latency sensitive network, and the rest over the high bandwidth network. The paper includes an algorithm to decide how to split traffic between the two networks.
“And the cool thing is that by accelerating a small fraction of the overall traffic for applications like web browsing – about 10% or less – can result in a huge benefit to the application.”
Godfrey explained that the success of this project became possible through a group diverse in their specializations and deep in terms of student involvement.
Student co-authors include Debopam Bhattacherjee, ETH Zürich; Waqar Aqeel, Duke University; Iker Nadi Bozkurt, Duke University; William Sentosa, Illinois CS; and Muhammad Tirmazi, Harvard University.
In addition to Godfrey, faculty collaborators included Sangeetha Abdu Jyothi, UC Irvine; Anthony Aguirre, UC Santa Cruz; Balakrishnan Chandrasekaran, VU Amsterdam; Gregory Laughlin, Yale University; Bruce Maggs, Duke University; and Ankit Singla, formerly of ETH Zurich and currently with Google. Both Abdu Jyothi and Singla are also Illinois CS alums.
“We knew that this was going to take multiple different kinds of expertise, including areas other than computer science,” Godfrey said. “It took a sustained effort by the team across multiple years.”
The project was funded by the National Science Foundation and a gift from Google.