Title: SIGCOMM Debates
Host: Akshay Narayan (Brown University)
Panelists: Ayush Mishra (NUS), Aditya Philip (CMU)
Scribe: Chengjin Zhou (Nankai University), Gao Han (Xiamen University)
Introduction
SIGCOMM should include a non-paper session centered on one or more moderated technical debates on specific networking-related topics. The SIGCOMM debate session will foucus on community deployability standards for congestion control algorithms.
Debate topic:
This debate centers on the trade-offs between fairness and performance in congestion control algorithms on the internet. The discussion explores the implications of adopting different approaches and establishing a deployability threshold that balances these aspects.
Aditya Philip’s Opening Statement:
Today, we’ll discuss the deployment of congestion control algorithms like BBR, which improve performance but often at the cost of fairness. Our focus is on finding a balance for deployability thresholds, considering both performance and fairness.
Fairness isn’t just about equal bandwidth. It’s crucial for usability. Imagine video calls where you can’t use the internet because other activities disrupt the service. That’s a bad outcome for everyone. We’re seeing new services like cloud gaming that need high throughput and low latency. If these services fail initially, they might not succeed. Fairness plays a vital role in determining what services flourish.
Deploying new algorithms often means testing them in different network settings. However, real-world testing on the internet can obscure whether performance gains come at the expense of fairness.
Ayush Mishra’s Opening Statement:
I argue that performance should be prioritized over fairness on the internet. The heterogeneity of congestion control on the internet is due to different goals across networks. Fairness is a legacy metric that doesn’t fully capture modern internet needs.
We’re debating fairness versus performance because we perceive them as opposites. But now, performance encompasses more factors. Isolation, as proposed by RCS, might be a better solution than fairness.
Enforcing fairness is practically challenging. Different congestion control algorithms have different goals, making it unfair to judge them by the same standards. Instead, we should focus on how to achieve isolation between competing goals to enable coexistence on the internet.
Host: Philip mentioned the “race to the bottom” idea. How do you respond to that concern, Ayush?
Ayush Mishra: The race to the bottom is due to a lack of isolation. For instance, BBR isn’t inherently unfair, but it’s incompatible with cubic, leading to issues. The focus should be on isolation rather than fairness. If we can isolate different strategies, we can avoid the race to the bottom.
Aditya Philip: CC developers can finally focus on improving performance because that’s what we set out to do. And let’s be honest, balance is hard and it makes a complicated problem even more complicated. I think maybe in the future, isolation would be a good goal being of. But in the present, fairness should still be an important part of when you decide to efficiently deploy your CC.
Host: What is an effective deployability criteria? Or does there exist one?
Aditya Philip: I guess the question is what, given that everyone cares about legacy fairness and everyone cares about throughput, we should most likely be able to converge on some deployability threshold that is in everyone’s interest. No one wants an internet that is high latency, and luckily, most large companies today have interest in both directions. So I think a deployability threshold that kind of takes industry stakeholders into the question, the equation should be able to come up with something that’s acceptable to everyone. And if it is acceptable to everyone, people should start using it to test it.
Ayush Mishra: I think I’ll come back to my initial point about these competing goals and different goals. And when you actually try to think of the deployability metric itself, it’s actually a really hard thing to come up with because different people have different goals and different objective functions.
Ayush Mishra’s Closing Statement:
As the internet evolves, we need to build infrastructure that accommodates diverse needs, like lanes for different vehicles on a highway. Isolation aligns with the internet’s heterogeneity, allowing different algorithms and services to coexist.
Aditya Philip’s Closing Statement:
Performance is crucial, but we can’t ignore fairness. It’s integral to the internet as a shared medium. We need to ensure that algorithms are fair and predictable before deployment to prevent issues down the line.
Personal thoughts
Different applications have different goals. There is no standard for whether a congestion control algorithm should focus more on performance or fairness. We should choose a congestion control algorithm based on the actual needs of the application. If this is a delay-sensitive application, focusing on performance is the best choice, but if an application requires user isolation, then fairness should take precedence over performance.