Grant: Discovery Grants Program - Individual

We are living in a time of exciting digital transformation and economic growth driven by cloud computing. Oracle predicted that 80% of enterprises plan to migrate their workloads to cloud data centers. However, Forbes reported that only 29% of engineers said their data centers can meet the current needs. Therefore, for next-generation cloud computing, the focus extends beyond the mere competition for high computational capabilities. Instead, "intelligent cloud services" are gaining rapid popularity, aiming to improve operational agility and efficiency without human intervention. This program focuses on designing new network architectures serving data-intensive applications in cloud data centers by leveraging, improving, and creating new Artificial Intelligence (AI) techniques to determine and optimize protocol mechanisms and control policies. First, we will explore how AI techniques can be applied and further improved in the network optimization domain, especially considering the heterogeneity and complexity of large-scale cloud data centers. The production systems rely on handcrafted heuristics, which are simple and easy to implement. However, the problem-specific heuristics rest on certain assumptions and predefined actions to certain events, which do not generalize well to changing environments. As a replacement for heuristics, existing clean-slate learning-based solutions have been proposed, which, however, raise new concerns such as high overhead, slow convergence, and bad performance over unforeseen conditions. In this program, we will combine traditional optimization strategies with machine learning and propose a set of practical and evolutionary learning-based solutions to achieve high throughput, low latency and overhead for next-generation cloud computing. Second, we will also examine how cloud data centers can be improved to better support AI applications in three dimensions: computation, communication, and network topology. As of today, AI training systems are predominantly built on top of traditional data center clusters due to their scalability, reliability, and cost-effectiveness. However, cloud computing was initially designed to achieve reliable resource sharing without optimization for specific applications. The network topology and the protocol stack used in cloud data centers need to be revolutionized to better support AI applications. In this program, we plan to co-optimize network topology, protocol, and parallelization strategy for accelerating the training of AI applications in cloud data centers. In summary, this program promotes the development of cloud computing, AI, and their application to the optimization of complex network systems. The proposed research will also contribute to the training of highly qualified personnel, leading to impactful research contributions and competitive job placements. Ultimately, the program's success will benefit the economy and society in Canada and around the world.