Streetlight Effect in Networking Research

Choosing which problem deserves our time is one of the hardest parts of doing research. Recently, Prof. Vyas Sekar’s HotNets paper^[1] made me pause and ask an uncomfortable question: are we working on the problems that matter most, or simply looking where the light is brightest?

The original paper

Original Figure 1 ^[1]: topics brighten and fade in SIGCOMM. Rows are topics; columns are years. Warmer colors mean larger paper shares.

The streetlight effect describes our tendency to search for answers where looking is easiest. In networking research, well-defined problems come with ready-made simulators, public traces, and standard benchmarks.^[1] Those tools make progress easy to measure and papers easier to defend during peer review. Over time, author, reviewer, and funding incentives reinforce this cycle. To break out of it, Prof. Sekar proposes building new streetlights through shared research infrastructure, outside perspectives, and broader evaluation criteria.

At the same time, the paper notes a second plausible explanation for why certain topics cluster on the heatmap. Our community may simply be adapting as real-world technologies and industry needs evolve. Topic counts alone illustrate what we publish, leaving the underlying cause open for debate.

Looking beyond titles and a single venue

Reading that paper made me curious about how the broader networking community behaves beyond SIGCOMM. A single flagship conference only captures part of our field, and a paper’s title rarely tells the full story. I extended the analysis across four major networking conferences: SIGCOMM, NSDI, IMC, and CoNEXT, covering all 4,154 research papers published through 2025. To keep the comparison fair, I removed short posters, demos, and keynotes, added a category for distributed systems and storage, and collected every paper’s full abstract.

Finding 1: Reading abstracts changes the topic labels for four in five papers. Prof. Sekar’s analysis was solely based on paper titles. Including abstracts updates the assigned topics for 82.2% of the papers and rescues hundreds of papers from the “Miscellaneous” bucket. This extra context even flips some recent trends. For example, at CoNEXT, judging by titles alone suggests that researchers apply AI methods to networking more often than they build networks for AI workloads. Reading the full abstracts reverses that ordering for 2021–2025 (15.2% on networking for AI workloads versus 12.0% on AI methods).

Analysis across 4 networking conferences

Finding 2: Each conference shines its own streetlight. Looking at SIGCOMM in isolation makes networking research appear more uniform than it actually is. Across the 2021–2025 window, each venue serves a distinct research agenda. NSDI leads in distributed systems and storage (28.0% of papers) as well as networking for AI workloads (16.6%). IMC anchors measurement research (86.1%) while expanding into security and privacy (56.9%) and web systems (46.2%). CoNEXT maintains the strongest focus on wireless and mobile networking (29.3%).

Conference SIGCOMMNSDIIMCCoNEXTAll four conferences Years 2005–2025Full conference history

Select a cell to inspect its papers below.

Conference agendas differ across venues. Rows are topics; columns are years. Darker blue means a larger share of papers; white means zero. A paper can match multiple topics. Select any cell to read its papers below, or choose “All four conferences” to compare them side by side.

Finding 3: Hot topics boom and fade across all four conferences at the same time. When a technology loses momentum, papers on that topic disappear across the entire community. Peer-to-peer and overlay networking accounted for 13% to 21% of papers across all four venues in 2005–2009, then dropped below 4% in every venue by 2021–2025. New hardware and workloads trigger the same synchronized surge. Over that same fifteen-year span, datacenter networking, programmable switches, and AI infrastructure climbed together across SIGCOMM, NSDI, and CoNEXT. Meanwhile, broad categories like measurement and security stay high decade after decade because researchers continually apply those methods to each new wave of systems.

Topic P2P & Overlay Years 2005–2025Full conference histories

Colored lines track each conference's three-year pooled paper share for the selected topic. Choose any topic and toggle conferences to compare their trajectories.

Finding 4: The research portfolio is getting broader, even at equal conference sizes. A natural worry about the streetlight effect is that our community might be converging on fewer topics over time. Because conferences accept many more papers today than they did twenty years ago, raw topic counts can easily mislead us. To compare years fairly, I sampled 19 papers per conference each year and measured how evenly those papers spread across topics. Across all four venues, the 2021–2025 portfolio is noticeably broader than the 2005–2009 portfolio. This broadening holds both under our 28 topic categories and when an unsupervised text model groups papers automatically by co-occurring words.

Measure Topic breadth at equal program sizesTopics inferred from paper text

Topic breadth rises even when conference sizes are held equal. “Effective topics” combines how many topics appear with how evenly papers are spread among them. A score of five has the same breadth as five equally common topics. All four venues have higher late-period averages, despite annual fluctuations.

What the numbers can and cannot tell us

Finding 5: Data cleaning and category boundaries can swing a heatmap cell by nearly 15 percentage points. Before reading too much into any publication chart, we have to check what enters the denominator. The SIGCOMM dataset from Prof. Sekar included 333 non-research records, mostly two-page posters and demos from 2010 through 2016. Removing those short entries while keeping the original topic labels fixed drops SIGCOMM’s 2010 wireless share by 14.5 percentage points and raises its 2015 datacenter share by 12.0 points. Category design shapes the visual impression just as strongly. Our taxonomy mixes broad methods like measurement with specific technologies like RDMA, which naturally makes broad categories look dominant.

Data cleaning changes many results. Bars show how 3 original SIGCOMM cells shift substantially after removing posters, demos, keynotes, and duplicates while holding topic labels fixed. Units are percentage points.

Finding 6: Popularity is no proof of misplaced effort. Even with complete abstracts across four venues, keyword rules remain an approximation because an abstract can mention a workload or protocol without making it the paper’s core contribution. More importantly, accepted papers only record what cleared peer review. Publication counts tell us nothing about rejected ideas, problem difficulty, or real-world value. A crowded topic like datacenter networking or AI infrastructure may attract heavy investment simply because that technology is reshaping computing. Conversely, a broader conference portfolio can still leave vital problems in the dark. Heatmaps can show us where the community has been, but deciding where your next unit of research effort matters most still requires your own judgment.

Papers used in the analysis

You can explore all 4,154 papers below by filtering by conference, topic, year, or keyword. Papers appear the newest first, with titles sorted alphabetically within each year.

Conference All conferencesSIGCOMMNSDIIMCCoNEXT Topic All topics Year Search

Loading paper evidence…

Conference	Year	Paper and abstract	Topics

References

1. [1] Vyas Sekar, “More Streetlights, Please: On the Streetlight Effect in Networking Research,” ACM HotNets, 2026.Back to citation 1 · Back to citation 2 · Back to citation 3