Projects

Understanding Traffic Bursts in GPU Collective Communication

Question. What execution mechanisms produce millisecond-scale traffic bursts during distributed AI training? I study how collective operations and nearby compute phases influence the network traffic observed at the NIC.

Approach. I align application-level collective timing with NIC traffic measurements and compare controlled communication-only versus compute-interleaved executions. The ongoing goal is to identify bottleneck mechanisms that can inform workload-aware scheduling, routing, and load balancing; I do not yet claim a deployed optimization system.

A Data Forwarding Scheme Using Proportional-Fair Scheduling in Edge Networks

Question. How can data forwarding adapt to differences in network-path quality under changing load?

Approach. I developed PFS-DFF, combining proportional-fair scheduling with data fragmentation and multipath forwarding based on relay load and link quality. I evaluated it using a Raspberry Pi and laptop wireless edge-network testbed. The thesis reports faster data reception, particularly under high load and for large transfers; my application materials report about a 9% latency reduction for a tested comparison.

Structural Dependency Modeling for Synthetic Data

Question. Can prior knowledge of relationships among features improve the fidelity of synthetic data?

Approach. I combined a Bayesian-network structure and conditional probability tables with CTGAN to generate synthetic electronic health records. On the UCI Diabetes 130-US Hospitals data, the hybrid method reduced pointwise mutual-information error by approximately 92% relative to CTGAN in the reported evaluation. This measures preservation of modeled dependencies; it does not by itself prove causal identification.

Learning Temporal Patterns in Network Traffic

Question. How much of traffic variation can be explained by its temporal history?

Approach. I built a Raspberry Pi-based LAN testbed, captured traffic traces, and compared LSTM-based traffic prediction with simpler regression baselines. This introduced me to measuring bursty traffic and studying the effects of temporal patterns using controlled experiments.

Multimodal Spatiotemporal Modeling

I worked on data analysis and preprocessing pipelines combining multiple data sources for urban crowd-density modeling, contributing to research on real-time high-resolution spatiotemporal reconstruction and prediction.