Date of Award
3-31-2024
Document Type
Thesis
Publisher
Santa Clara : Santa Clara University, 2024
Degree Name
Master of Science (MS)
Department
Computer Science and Engineering
First Advisor
Behnam Dezfouli
Abstract
Wi-Fi is the primary wireless communication method for the majority of devices in both residential and commercial settings. The number of devices continues to increase, making latency, bandwidth, and security difficult to manage and balance to provide satisfactory performance for all users. Novel methods that rely on the collection of data from the networking stack have been developed in research settings to analyze and predict key parameters and network state to improve communication efficiency. However, crucial processes such as experimentation, data collection, and performance analysis have often been performed manually on offline data.
This thesis presents GigaCollector, a scalable end-to-end framework to enable temporally-coherent high-rate, realtime, and flexible environment control, experiment control, data collection, data collation, and model inference for the research, experimentation, and development of new machine learning-based edge systems such as Wi-Fi schedulers for improved latency and power consumption. It is also applicable to other fields such as sensor fusion for motion capture. The framework consists of five main components: (1) Active Environment Control (AEC), (2) the Experiment Controller (EC), (3) Data Sources (DSs), (4) the Collector & Collator (CC), and (5) Snapshot Consumers (SCs). Each component uses various techniques to enable the three core requirements: temporal coherence, high-rate and real-time operation, and flexibility. The AEC component uses closed-loop control algorithms incorporating system feedback to adaptively control experimental conditions. The EC automates AEC configuration, iterates through user-defined test cases with programmable start and stop conditions, and records all results. The DS, CC, and SC interfaces use ZeroMQ, a lightweight, low-latency, high-throughput, and open-source messaging system to enable language-agnostic interoperability between local and remote distributed components. The CC uses a Temporal Index-Matched List (TIML) data structure to allow large data history storage with O(log(n))-class fuzzy closest-timestamp data collation for temporally-coherent snapshots.
The results show that this framework is able to achieve its two core goals of (1) flexible, scalable, temporally-coherent, high-rate, and real-time data collection and snapshot collation, and (2) real-time closed-loop experimental environment control. For goal (1), on a 24-core dual-Xeon E5-2690 v3 workstation, GigaCollector achieves sub-millisecond end-to-end latency using a single processor core with an average of 500 microseconds (300 microseconds for collection, 70 microseconds to process, store, and collate data across 10 data sources each with 10 elements, and 150 microseconds to deliver the data to a snapshot consumer) with an average maximum message processing rate of 13,000 per second per CC instance under reasonable loads. Horizontal scalability with two parallel instances doubles throughput in certain circumstances with minimal impact on end-to-end latency. For goal (2), the system achieves independent control of uplink and downlink airtime utilization with roughly 10% deviation from the set point using an Atheros AR9462 Wi-Fi network interface connected to the same workstation as the server and a Raspberry Pi 5 as the client.
Recommended Citation
Vu, An, "GigaCollector: A Real-Time, Temporally Coherent Framework forWi-Fi Environment Control, Data Collection, and ML Inference" (2024). Computer Science and Engineering Master's Theses. 44.
https://scholarcommons.scu.edu/cseng_mstr/44
