Trying to create a network that is fair, equitable, and starvation-free may simply be impossible, at least with current congestion control algorithms (CCA), an MIT study has found.
Published ahead of a conference presentation this week, the MIT paper [PDF] found that regardless of approach, CCAs like Google’s BBR, FAST, and others suffer from the physical limitations of networks, leading to an unavoidable problem where some users lack bandwidth.
“Our theorem shows that CCAs have to choose at most two of three properties: high throughput, convergence to a small and narrow delay range, and no starvation,” the researchers said in the paper.
The paper cites a number of non-congestive network issues, such as ACK aggregation and end-host scheduling, that serve to override strict algorithmic controls that rely on guesswork to account for events in a network that it cannot control. .
In an ideal situation, the researchers wrote, CCAs operating on a single network are designed to converge and work together to create the smallest delay range possible. According to the researchers, this is where the problem lies.
“Because most CCAs try to work over many orders of magnitude of rates, they must map a large range of rates onto a small range of delay. Thus, even small changes in the estimated queue delay would induce huge changes” the team wrote.
In other words, the algorithms account for a lot, but simply cannot account for real-world physical imperfections or non-congestion-related delays in their calculations.
Can we build better CCAs?
The paper admits that its conclusions are “pessimistic for CECs that limit delays” and raises the question of “whether we are doomed to choose between limiting delays and avoiding starvation.”
speaking to IEEE Spectrumlead study author and MIT computer scientist Venkat Arun said his team’s discovery sheds new light on CCA problems previously attributed to poor algorithmic decision-making and insufficient network capacity.
Instead, Arun and his team’s research shows that the algorithms themselves are simply not designed to account for network fluctuations, which the paper uses to refer to non-congestive causes of network delays. “We do not think it is possible to circumvent this problem with algorithms that map loss rates (or delays) to shipping rates,” the team wrote.
This is not to say that the MIT team doesn’t have suggestions on how to solve this seemingly unavoidable network management problem. In the document, the team makes several suggestions that basically amount to increasing algorithmic queue times to account for nerves.
Still, even that might not be enough, the team concluded. “It’s also possible that purely end-to-end CCAs will always suffer from the problems we found, and support is required in the network, such as active queuing, explicit congestion signaling, or stronger isolation.” ®