Systems and Networking Lab, UC Santa BarbaraPramana
Try the demo

HotNets 2026 · Salt Lake City

Pramana

A Composable, Domain-Specific Backend for Empirical Networking Research

PaperVideoAnnotated deckCodeBibTeXInteractive demo
Pramana as a thin waist: research intents (what) at the top, substrates such as local, cloud and testbeds (where) at the bottom, and Pramana composing mechanisms (how) in the middle, returning labeled data.
Pramana is a thin waist for networking research. A researcher states what data a study needs (the intent); Pramana picks how to produce it (the mechanism) and where to run it (the substrate), then returns labeled data.

Try three intents from published papers

Pick an experiment from a published paper. Pramana checks the intent, writes the spec, composes the workflows, runs them, and plots what comes back.

Open the demo
SIGCOMM '24CCAnalyzerIllustrativeWhich congestion control does an Azure server run?Downloads 100 MB over 4 bottleneck settings, records queue occupancy, and classifies it with 1NN-DTW.Illustrative verdict: BBR (3 of 4 settings agree)The paper regenerates CCAnalyzer's own traces, one trial per setting and algorithm. This Azure run and its verdict are illustrative.SIGCOMM '24PrudentiaIllustrativeDoes a Google Meet call keep its quality next to an iPerf Cubic flow?Walks through the paper's real-time experiment on a shared bottleneck, with resolution, frame rate, freezes, and delay as the metrics.Illustrative: Meet quality under Cubic competitionNot yet run: the implementation has not run the Meet call and the iperf3 download together, and reports no call metrics (§3).Rejection pathStarlink / LEODesign behaviorHow does a Zoom call behave over a Starlink access link?Pramana maps the intent, then checks each part of it against what the substrates offer.Rejected: none of SNL, PINOT, or AWS has a Starlink dishThis is the design's check. It needs a capability file from each service, and none publishes one yet (§3).

Abstract

Empirical networking research often tests a hypothesis with data the study itself generates, so accelerating this research requires shortening the time from forming a hypothesis to generating that data. Consider whether a bulk BBR download fairly shares its bottleneck with a Google Meet call. Testing this requires configuring a realistic bottleneck link, running the BBR transfer and the Meet call together, and collecting the call's quality metrics. Today a researcher builds this setup from scratch for each new idea; this delay from idea to data is the ideation-to-data gap. AI agents already propose hypotheses faster than researchers can generate data to test them. Closing the gap needs a backend that generates the data for each new hypothesis on demand, without a new setup.

We envision Pramana, a composable backend for networking research designed as a thin waist, with research intents above it and execution substrates below. At its core, the intent specification separates the data a study needs (the intent) from where that data is generated (the substrate) and how it is produced (the mechanism), so one specification can target any substrate. We build a corpus of 255 data-generation intents from 66 published papers and find that the intent specification expresses all of them, while no existing tool covers more than 11%. Our proof-of-concept implementation covers 27% of these intents. For CCAnalyzer, one of the 66 papers, the implementation regenerated the data for one trial of each network setting and congestion-control algorithm in the paper's main experiment. We lay out what implementing the full design requires and why it needs a community effort.

Coverage of 255 intents from 66 papers

data-generation intents in the corpus
255
published papers they come from
66
expressible in the intent specification
100%
covered by the proof-of-concept implementation
27%
Share of the 255 corpus intents each system satisfies
Intent specification100%
Proof-of-concept implementation27%
Best existing tool (NetReplica)11%

Video

HotNets '26 talk · 2:41

Pramana's architecture: a client, parser, executor, data store and telemetry, calling external services (an LLM, Mininet, NetGent, NetReplica) and deploying to local, cloud or testbed substrates.

How it works

WHAT

Intent

The data a study needs: the apps, the traffic, the metrics to record.

WHERE

Substrate

Where the data is generated: a laptop, a cloud region, or a testbed like PINOT or FABRIC.

HOW

Mechanism

How it gets produced. The implementation composes two existing systems, NetReplica and NetGent.

BibTeX

@inproceedings{daneshamooz2026pramana,
  title     = {Pramana: A Composable, Domain-Specific Backend
               for Empirical Networking Research},
  author    = {Daneshamooz, Jaber and Vuong, Eugene and Ramanathan, Alagappan
               and Moghimi, Manni and Manda, Haarika and Kumar, Satyam
               and Thode, Snithik and Guthula, Satyandra and Beltiukov, Sylee
               and Han, Dongsu and Mangla, Tarun and Abdu Jyothi, Sangeetha
               and Willinger, Walter and Gupta, Arpit},
  booktitle = {Proceedings of the 25th ACM Workshop on Hot Topics in Networks},
  series    = {HotNets '26},
  address   = {Salt Lake City, UT, USA},
  year      = {2026}
}