Research Projects

Cloud storage services have experienced exponential growth due to the increased usage of platforms like Dropbox, OneDrive, and Google Drive, which offer ease of data access and management. Similarly, real-time multimedia traffic has surged, driven by the rise of online meetings and lectures, which have become the new norm post-COVID. This significant increase in data storage and network throughput requirements necessitates innovative methods to manage and optimize these resources effectively. The goal of this proposal is to explore two broad directions to tackle these requirements. (1) The first direction is theoretical, aiming to mathematically model the requirements and study the fundamental limits of erasure codes for the distributed storage systems by applying techniques from coding theory. (2) The second direction is experimental, focusing on systems design of latency-sensitive real-time streaming applications such as video-call services and virtual and augmented reality (VR and AR) use cases. Design of codes for distributed codes has been initially focused on coming up with MDS codes with minimal repair bandwidth. However, these codes are limited by high sub-packetization level to realize the theoretical gains they offer in actual systems implementation. This led to study of near-optimal MDS codes that are MDS codes with fixed sub-packetization but have repair bandwidth slightly higher than that of the optimal repair bandwidth. Though there exist frameworks to construct near-optimal MDS codes, the fundamental lower bounds on the repair bandwidth required for a fixed sub-packetization level are not known and the best known near-optimal MDS codes in terms of repair bandwidth are not explicit constructions i.e, they require large field size. As part of this project, we plan to therefore study these fundamental limits and also design codes with low field size, better disk access properties for Reed Solomon [14, 10] and [9,6] codes used in practice. As part of the second direction, we plan to set up an end-to-end video streaming simulation system that offers flexibility in employing various streaming code options available in the latest literature. This system will be evaluated using open-source network traces of video traffic. Additionally, we plan to use machine learning techniques to develop adaptation algorithms that can determine the optimal compression ratio and streaming code parameters based on network statistics such as packet loss ratio and available throughput.