People of ACM - Sylvia Ratnasamy

April 9, 2026

Your 2001 paper “A Scalable, Content-Addressable Network,” is considered one of the most influential recent works in computer networking. The paper examined how to scale a peer-to-peer (fully decentralized) file sharing network (such as Napster). What was the key innovation you presented in this paper? Given that centralized streaming has become dominant (with apps such as Google Drive and Dropbox), what is the legacy of your work with decentralized content addressable networks?

 The key innovations were twofold. The first was the concept of a distributed hash table that provided a simple hash-table interface—put (key, value) and get (key)—at a global Internet scale. The second was the decentralized design to implement this interface, which involved partitioning a d-dimensional Cartesian coordinate space across a set of participant nodes. A node was responsible for a portion of the space and connected only to its neighbors in the space. To find an object, we’d hash the object name to a point in the coordinate space and then route through this virtual space until we found the node responsible for that point.

I’d say that the legacy of DHTs—and here I’m referring to the body of DHT research and not just my own papers—was primarily the DHT abstraction rather than the specific routing design. The DHT literature demonstrated that a massive array of services—file systems, global name resolution, search, and distributed storage—could be built on this abstraction, and we see the DNA of this DHT concept in modern distributed database and cloud storage services. The routing design itself is less adopted because complete decentralization is rarely required in modern services and adopting some degree of centralization often simplifies the system. Nonetheless, requirements evolve and I like to think that the DHT literature gives us the 'break glass in case of emergency' toolkit if full decentralization becomes a requirement again.

Your work with RouteBricks, as outlined in a 2011 paper you co-authored, introduced a new approach that allowed consumer-based routers to reach network speeds of 35-40 Gigabytes per second (Gbps). At the time, general-purpose routers were only reaching about 1-5 Gbps. What was the key innovation in RouteBricks that allowed routers to scale, while making it easier for networks to build, program and evolve?

For context, I did this work while at Intel Research. At the time, the transition to multicore CPUs was a paradigm shift that was causing programmers to fundamentally rethink how they wrote software, and it was clear that smart parallelization was the name of the game. Our key innovation in RouteBricks was the parallelization strategy that enabled networking workloads to fully exploit the capabilities of multicore hardware. These techniques—dealing with how we leverage multi-queue NIC hardware, thread pinning, NIC polling, cache partitioning, etc.—provided an architectural blueprint that is now standard practice in high-performance network stacks and network function virtualization (NFV) products. We showed that, with careful parallelization, a commodity server could scale to a traffic processing line rate of 10Gbps (which was a leapfrog advance at the time). These results triggered multiple follow-on efforts, especially in the context of kernel-bypass stacks such as Intel’s DPDK. They also helped mark out a role for software-based traffic processing as a valuable companion to traditional ASIC-based traffic processing.

Some experts predict that 6G technology, prevalent by the early 2030’s, may reach data speeds of up to 1 Terabit per second (1 Tbps). These speeds would be 10,000 times faster than today’s 5G. Do you agree with these predictions? What infrastructure improvements will be needed to reach these milestones?

As a systems researcher, my work begins where the radio layer ends, so I’ll leave the radio and signal layer questions to my EE colleagues. However, even if 1 Tbps is achievable as a peak physical-layer rate, I am skeptical that we will see those speeds reflected in the end-to-end user experience anytime soon. There is a big gulf between the laboratory peak rate and the end-to-end application throughput that users enjoy. To make 1 Tbps a reality that is available to the general public, we would have to radically rethink the entire architectural pipeline. For example, we would require orders-of-magnitude increase in cell-site density. In addition, there is the backhaul infrastructure. It does no good to have 1 Tbps per user if the fiber infrastructure connected to the tower cannot relay that traffic onwards. Finally, we would have to reconsider the end host systems. Our current mobile stacks and processors aren't designed to sink or source 1 Tbps of data. Driving that kind of throughput from a smartphone app would require a fundamental redesign of the networking stack and mobile hardware. In short, the systems challenges here appear just as daunting as the physics. But I’d love to be proven wrong!

What research project are you dedicating most of your time to now?

Lately my work has focused on the resilience of large-scale infrastructures—global-scale networks, datacenter clusters, and cellular networks. These systems are critical to modern life, but they also—perhaps necessarily—have become incredibly complex. Despite massive effort and investment, we still see these infrastructures fail in troubling ways. Recent outages have disrupted everything from cloud services and 911 calling to widespread blackouts that resulted from faulty software upgrades.

These outages are disruptive to society and stifle innovation by making operators understandably conservative about change. More fundamentally, these outages highlight a gap in our field: while our community has built impressively capable infrastructures, we have fallen short when it comes to building systems whose performance and behavior we can deeply reason about. This is a fundamental flaw in our discipline. As a systems researcher, this bothers me and I want to help develop solutions where “performance clarity” is a first-class design principle rather than an add-on.

There is a vast amount of work that remains here, even more so as AI emerges as both a demanding workload and a powerful new tool. My current efforts are focused on these questions: how do we architect systems so that they’re easier to reason about; how do we leverage AI to predict and diagnose problems; how do we generalize our results to different infrastructures.

In a recent OpEd in Communications of the ACM titled “Resilient Infrastructures via Digital Unification” you and your co-authors envision a “digital unification of industrial architectures.” Will you explain what is meant by this term and why it is needed?

Our group, led by Ang Chen, wrote the OpEd as a call to action. It’s clear that the digital transformation of infrastructures like power grids or water systems (where software is embedded into their operations) is underway. The problem is that we’re seeing it happen in silos. Currently, each sector develops point solutions with little cross-industry discussion or broad engagement from the systems and networking community.

Our core observation is that regardless of the target infrastructure, there are a lot of commonalities in the underlying systems requirements. Increasingly we’re seeing common design patterns—like data-driven control loops that collect telemetry and configure or actuate devices. At the same time, common goals and requirements around scalability, security, resilience, programmability, are coming to the fore.

We are thus calling for digital unification, by which we mean developing common design principles and abstractions rather than fragmented solutions. Note that we aren’t suggesting that all infrastructures run a single computing stack. That would be a bad idea. Rather, we’re talking about a shared foundation of concepts, abstractions, and design blueprints much like the Internet’s TCP/IP protocols or the POSIX abstractions in operating systems that provide a common language for application developers. Unification offers several practical benefits. It allows traditional infrastructures to leapfrog to state-of-the-art computing solutions. Richer inter-infrastructure coordination is also being enabled. (e.g., datacenters and the grid sharing information to optimize energy use for the AI era). Finally, it is improving resilience by hardening a smaller consolidated set of building blocks.

Identifying these modular, general abstractions is what systems researchers do best and have successfully done for the Internet and modern computing. Our paper invites the community to apply that same mindset to a broader set of physical infrastructures that are both critical to society and eager for transformation.