Distributed Denial of Service (DDoS) continues to be a significant threat to the availability and stability of networked systems, including smart city ICT infrastructure. This research project engages students in exploring and evaluating machine learning (ML) approaches for DDoS detection using the NSF FABRIC testbed.

Students will build virtual network environments, launch attacks mixed with benign traffic, collect network flow data, and train ML/DL models to evaluate attack classification performance. Advanced students will have the opportunity to develop adaptive, online detection methods for new and evolving attacks.

Student Outcomes: Students will gain hands-on experience in network flow analysis, ML applications, and cyberinfrastructure experimentation, while contributing to innovative solutions in network intrusion detection.

Classical computing and communication is an integral part of the pipeline of quantum computing. Machine learning (ML) and AI models can play a significant role in improving, accelerating, and scaling quantum computing platform especially the quantum error correction codes, critical to quantum computing development.

This research project engages students in researching quantum computing by conducting simulation studies and prototyping new ideas using the distributed computing resources, especially the GPUs, in the NSF FABRIC testbed. 

Students will explore the quantum error correction codes (QEC), use existing simulation packages, and build a basic simulation pipeline on the AI computing cluster to study the performance of representative QEC performance under realistic hardware and networking constraints.

Student Outcomes: Students will gain theoretical and hands-on experience in quantum computing and fault tolerance, as well as using modern cyberinfrastructure. This research may lead to innovative QEC solutions.

Programming Protocol-independent Packet Processors (P4) is a cutting-edge language for network data plane programming, enabling researchers—not just device vendors—to define and program network-connected devices such as switches and smart-NICs. This project introduces students to deploying P4 applications on FPGAs using tools and resources in the NSF FABRIC testbed.

Students will explore advanced networking topics, including In-band Network Telemetry (INT), network security, and networking support for machine learning. Example objectives include developing an INT-based approach for one-way delay measurements using FABRIC’s GPS-synchronized time signals and P4-programmable FPGAs, followed by large-scale evaluation across distributed FABRIC sites.

Student Outcomes: Participants will gain hands-on experience developing and deploying P4 applications, conducting advanced networking experiments across FABRIC sites, and building INT solutions that can be leveraged by the broader research community.

This project involves implementing a P4-based packet processing pipeline on the NVIDIA BlueField-3 DPU. The application will profile network traffic destined for the host by performing real-time flow classification in the data plane. Specifically, it will identify and count the number of active flows using common protocols such as HTTP, HTTPS, DNS, as well as other traffic types. In addition, the project includes exporting traffic statistics from the data plane to the control plane, where the collected metrics will be visualized using a Grafana dashboard. 

Student Outcomes: The student will learn real-time monitoring and analysis of host-bound network traffic with minimal CPU overhead. The student will gain experience in tracking per-flow statistics and analyzing metrics at high rates (100 Gbps). The student will gain hands-on experience developing and deploying P4 applications, and conducting advanced networking experiments across FABRIC sites.

TCP congestion control algorithms (CCAs) are essential for ensuring efficient data transmission, yet there is no single “best” approach. Traditional CCAs like CUBIC and Reno use Additive Increase Multiplicative Decrease (AIMD), while newer algorithms like BBR estimate round-trip time and bottleneck bandwidth to optimize sending rates.

In this project, students will evaluate and compare different CCAs on the NSF FABRIC testbed, focusing on their performance in large-scale scientific data transfers. Participants will design experiments, analyze fairness and interactions among CCAs, and study their performance in networks with security appliances such as firewalls and IDS/IPS.

Student Outcomes: REU students gain hands-on experience in network performance analysis and congestion control, developing practical skills in deploying and testing CCAs on real testbeds. They will also create interactive labs and Jupyter notebooks to share insights with the broader FABRIC community.

Introduced by Google in 2016, the first version of the Bottleneck Bandwidth and Round-Trip Time (BBRv1) congestion control algorithm (CCA) marked a significant advancement in network communication. Unlike traditional loss-based CCAs such as CUBIC and Reno, BBRv1 focused on balancing throughput and delay without relying on packet losses as a congestion signal. However, BBRv1 faced fairness issues due to its aggressiveness when interacting with loss-based CCAs. To address this issue, BBRv2 was developed, incorporating multiple metrics to improve fairness with loss based CCAs. Despite improvements, BBRv2 encountered bugs and performance limitations, prompting the release of BBRv3 in 2023. BBRv3 addressed bugs found in BBRv2 and fine-tuned parameters to enhance flow coexistence.

This project will study the performance of BBRv3 across diverse network scenarios by comparing it with CUBIC and further contrasts this comparison with those involving CUBIC against BBRv2 and BBRv1. The project will explore BBRv3’s behavior under different conditions, considering variations in the number of flows, propagation delays, loss rates, and buffer sizes. Additionally, the project will study the impact of Active Queue Management (AQM) algorithms in addressing the RTT unfairness issue.

Student Outcomes: Students will gain hands-on experience in network performance analysis and congestion control, developing practical skills in deploying and testing CCAs on real testbeds. They will also create interactive labs and Jupyter notebooks to share insights with the broader FABRIC community.

A network testbed requires substantial observability for its users and operators to capture transient conditions of the testbed that might be difficult to reproduce, and that affect the testbed’s performance of user experiments. To provide observability into FABRIC’s network, Patchwork is an award-winning network profiler for FABRIC that programmatically samples the network to create a trace of its contents over time. Building a testbed-wide profiler is non-trivial since it involves making various kinds of technical trade-offs, therefore Patchwork is designed to accommodate several choices to balance between utility and overhead. This design offers several dimensions along which REU students can add features and help explore the consequences of different technical choices. For these REU projects, we envisage Patchwork serving as a platform on which a student will build and evaluate extensions for custom measurement and analysis of network traffic.

The goal of these student projects is to build and evaluate integration between Patchwork and third-party, widely-used network monitoring approaches. This research seeks to understand (1) how to build and evaluate such integrations, and (2) form an initial understanding of how FABRIC users can obtain more detailed insights about their experiments by using Patchwork-supported monitors. This research builds on Patchwork’s programmability to explore different approaches for implementing the integration, and for fine-grained measurements to evaluate each approach.

Project 1: Integrating analysis tools

This project seeks to integrate NetFlow tools or GraphBLAS with Patchwork, to feed Patchwork-captured traffic to those tools. Such tools have already been packaged to run on FABRIC during preliminary work by IIT students, but they have not yet been integrated with Patchwork. By integrating such a tool, the student will research a pathway for other researchers to use third-party analysis tools with Patchwork for their experiments.

Project 2: Adding custom analyses

Currently the Patchwork dashboard does not show flow-level information of network traffic.Flow-level information provides granular insights into the characteristics of network traffic. While this analysis was carried out as part of the Patchwork paper, it is not yet integrated with the dashboard. The goal of this project is to port over this analysis to be included in the dashboard. Then, using that integration, the student will have a foundation on which they can research different types of flow-level analyses and explore trade-offs.

Student Outcomes. As a result of these projects, the student will learn about the challenges and techniques related to network monitoring, and will gain skills for diagnosing, measuring, and debugging networks and their monitors. Students will also contribute to Jupyter notebooks and will be encouraged to present or write up their work. These notebooks will be open-sourced for the benefit of the FABRIC community, and will serve to demonstrate the diagnosis, measurement, and debugging of distributed systems over FABRIC. Students will be mentored to prepare a demo about their work, and they will also be invited to integrate their work for display in Patchwork’s dashboard.

AI-assisted medical imaging combines networking, image processing, machine learning, and visualization to enable real-time analysis of medical images for clinical decision-making. This research addresses challenges such as high-quality image acquisition, optimized data transfer, algorithm development, and clinically relevant visualization.

Students in this project area will use the NSF FABRIC testbed—along with resources from Chameleon and a high-resolution Philips ultrasound system at Clemson University—to provision, configure, and deploy experiments in AI-assisted medical imaging. Students will explore end-to-end systems, including software, compute, and networking resources, and may investigate machine learning performance or image visualization methods.

Project 1:Machine learning across FABRIC and Chameleon  

In this project, the student will learn to use FABRIC Artifact Manager to create experiments across FABRIC and Chameleon, retrieve data from a remote archive and move them to Chameleon GPU nodes, and perform machine learning on them.  Students can either choose their own machine learning example or be guided with our medical imaging example.   

Project 2: Secure streaming of medical ultrasound video and images across SciStream 

 In this project, the student will learn to use FABRIC Artifact Manager to create an experiment with SciStream secure transport between two FABRIC nodes and then explore adaptive streaming solutions on it.  Students will explore available adaptive streaming software online.  In addition, students can explore advanced compression techniques.

Student Outcomes: REU students will gain hands-on experience designing and running multi-site FABRIC experiments, interfacing with external testbeds, measuring network performance, and working with medical imaging data, preparing them for research across multiple disciplines.