The growing demand for efficient communication in distributed training and inference has sparked significant interest in reconfigurable photonic interconnects across both academia and industry. Mixture-of-Experts (MoE) models, with their highly skewed communication patterns, present a natural opportunity for such circuit-switched fabrics. However, existing approaches largely optimize communication in isolation, overlooking the interaction between communication and the expert computation that follows.

AllReduce is a fundamental collective communication operation in distributed computing and a key performance bottleneck for large-scale training and inference. Its completion time is determined by the number of communication steps, which dominate latency-sensitive workloads, and the communication distance affecting both latency- and bandwidth-bound regimes. Direct-connect topologies, such as Google's TPUv4 tori, are particularly prone to large communication distances due to limited bisection bandwidth.

As chip-to-chip silicon photonics gain traction for their bandwidth and energy efficiency, their circuit-switched nature raises a fundamental question for collective communication: when and how should the interconnect be reconfigured to realize these benefits? Establishing direct optical paths can reduce congestion and propagation delay; however, each reconfiguration incurs non-negligible overhead, making naive per-step reconfiguration impractical.

In modern datacenters, the effectiveness of end-to-end congestion control (CC) is quickly diminishing with the rapid bandwidth evolution. Per-hop flow control (FC) can react to congestion more promptly. However, a coarse-grained FC can result in Head-Of-Line (HOL) blocking. A fine-grained, per-flow FC can eliminate HOL blocking caused by flow control, however, it does not scale well. This paper presents Pyrrha, a scalable flow control approach that provably eliminates HOL blocking while using a minimum number of queues. In Pyrrha, flow control first takes effect on the root of the congestion, i.e., the port where congestion occurs. And then flows are controlled according to their contributed congestion roots. A prototype of Pyrrha is implemented on Tofino2 switches. Compared with state-of-the-art approaches, the average FCT of uncongested flows is reduced by 42%-98%, and 99th-tail latency can be$1.6\times $-$215\times $lower, without compromising the performance of congested flows.

As chip-to-chip silicon photonics gain traction for their bandwidth and energy efficiency, collective communication has emerged as a critical bottleneck in scale-up systems. Programmable photonic interconnects offer a promising path forward: by dynamically reconfiguring the fabric, they can establish direct, high-bandwidth optical paths between communicating endpoints — synchronously and guided by the structure of collective operations (e.g., AllReduce). However, realizing this vision — when light bends to the collective will — requires navigating a fundamental trade-off between reconfiguration delay and the performance gains of adaptive topologies.

We introduce self-adjusting partially ordered lists, a generalization of self-adjusting lists where additionally there may be constraints for the relative order of some nodes in the list. The lists self-adjust to improve performance while serving input sequences exhibiting favorable properties, such as locality of reference, but the constraints must be respected.We design a deterministic adjusting algorithm that operates without any assumptions about the input distribution and without maintaining frequency statistics or timestamps. Despite the more general model, we show that our deterministic algorithm performs closely to optimum (it is 4-competitive). In addition, we design a family of randomized algorithms with improved competitive ratios, handling also a more general rearrangement cost model, scaled by an arbitrary constant d ≥1. Moreover, we observe that different constraints influence the competitiveness of online algorithms, and we shed light on this aspect with a lower bound.We investigate the applicability of our self-adjusting lists in the context of network packet classification. Our evaluations show that our classifier performs similarly to a static list for low-locality traffic, but significantly outperforms Efficuts (by factor 7x), CutSplit (3.6x) and the static list (14x) for high locality and small rulesets.

Today's network devices share buffer across queues to avoid drops during transient congestion and absorb bursts. As the buffer-per-bandwidth-unit in datacenter decreases, the need for optimal buffer utilization becomes more pressing. Typical devices use a hierarchical packet admission control scheme: First, a Buffer Management (BM) scheme decides the maximum length per queue at the device level and then an Active Queue Management (AQM) scheme decides which packets will be admitted at the queue level. Unfortunately, the lack of cooperation between the two control schemes leads to (i) harmful interference across queues, due to the lack of isolation; (ii) increased queueing delay, due to the obliviousness to the per-queue drain time; and (iii) thus unpredictable burst tolerance. To overcome these limitations, we propose ABM, Active Buffer Management which incorporates insights from both BM and AQM. Concretely, ABM accounts for both total buffer occupancy (typically used by BM) and queue drain time (typically used by AQM). We analytically prove that ABM provides isolation, bounded buffer drain time and achieves predictable burst tolerance without sacrificing throughput. We empirically find that ABM improves the 99th percentile FCT for short flows by up to 94% compared to the state-of-the-art buffer management. We further show that ABM improves the performance of advanced datacenter transport protocols in terms of FCT by up to 76% compared to DCTCP, TIMELY and PowerTCP under bursty workloads even at moderate load conditions.