Datacenters are increasingly designed to support AI training workloads, which generate heavy, bursty traffic that leads to packet drops and severely degrades Collective Completion Times (CCTs). Existing transport protocols struggle to manage this congestion efficiently. In this talk, I introduce dcSIM, a novel transport protocol that utilizes a "Simulate Before Sending" framework on commodity switches. By sending small simulation probes to explore candidate paths and reserve buffer space, dcSIM ensures that data packets are only transmitted when a path is verified to be clear. Finally, I will demonstrate how dcSIM eliminates data packet collisions to achieve zero packet loss and maintains ~97% network utilization. As a result, it lowers CCTs significantly and remains highly resilient even in challenging oversubscribed topologies.
M.Sc. student under the supervision of Prof. Isaac Keslassy.