Dipartimento di Informatica

Abstract: The design of efficient interconnection networks is critical to the performance of modern large-scale computing systems. In this talk, I will present my research on three key aspects of this challenge: the design of network topologies, the optimization of collective communication algorithms, and the use of in-network compute to accelerate communication. These areas are essential for enabling scalable performance across a variety of applications, with distributed deep learning being one prominent example. I will discuss how architectural and algorithmic innovations can help overcome communication bottlenecks, highlighting insights from recent work and outlining opportunities for future improvements.