Congestion control in hyperscale datacenters underpins nearly every service, enabling the use of the network infrastructure while keeping communication among servers orderly. As workload requirements diversifyâ e.g., throughput-intensive ML training and disaggregated storage, and latency-critical microservicesâ link bandwidth keeps growing, but relative switch buffer capacity is not, reducing headroom to absorb traffic bursts. Against this backdrop, the question is how congestion control protocols can maximize bandwidth utilization without inducing latency- and loss-amplifying network queues.

In this context, receiver-driven transports, which explicitly schedule transmissions rather than react to congestion signals, excel at dealing with congestion on the last hop, because that hop is managed by a single receiver. However, when multiple receivers have to share a link (e.g., when they share a senderâ s uplink), their independent schedules can conflict.

This thesis argues that single-owner links should be proactively scheduled by receivers, while shared links should be governed by a reactive control algorithm that exposes their instantaneous capacity. The approach allows receivers to both precisely schedule their own links and coordinate over shared bottlenecks, which critically include sender uplinks. To realize this design, this thesis contributes SIRD, a Sender-Informed, Receiver-Driven transport. SIRD implements receiver-driven scheduling that incorporates congestion information from shared links in the form of (i) ECN from the networkâ s core, and (ii) end-host signals from senders. The result is tight scheduling, which enables high bandwidth utilization with little contention, and thus minimal latency-inducing buffering in the fabric. This thesis also contributes two enhanced versions of SIRD: (i) SIRD-RSF, which incorporates rich sender feedback to ac- celerate convergence, and (ii) SIRD-PrioFlow, which aligns reactive control with the desired scheduling policy to reduce latency.

Using large-scale simulations, we show that SIRD simultaneously maximizes link utilization, minimizes switch buffering, and attains near-optimal latency. Additionally, SIRD-RSF reduces convergence time by at least 2Ã and SIRD-PrioFlow reduces 99th-percentile latency for medium-sized messages by up to 50%. We also implement SIRD on a kernel-bypass software datapath and show that it is able to incorporate congestion feedback and tightly schedule transmissions at 100 Gbps.