Aug 2026· 2026 IEEE 51st Conference on Local Computer Networks (LCN)· pp. 1-12· 0 citations· 36 references
Computer Science
TL;DR
Probabilistic state-proportional (PSP) dispatching, a packet-level load balancing algorithm using a Band-based discrete state representation, provides an effective balance among performance, stability, and overhead for artificial intelligence data centers.
Abstract
With the rapid growth of large language model training and generative artificial intelligence services, data center networks face severe micro-burst traffic and high concurrency. Traditional hash-based flow-level load balancing cannot sense link states, leading to hash collisions, hotspot congestion, and tail latency in multipath Clos networks. Existing packet-level schemes are constrained by stale state information, high hardware complexity, and poor adaptation to heterogeneous links. To address these issues, this paper proposes probabilistic state-proportional (PSP) dispatching, a packet-level load balancing algorithm. Using a Band-based discrete state representation, PSP replaces global sorting with local probability mapping, reducing hardware complexity while suppressing herding and oscillations caused by stale states. Experiments on a cycle-accurate simulator show that PSP is robust across port scales, bandwidth-limited paths, and fixed-flow interference. It outperforms join-the-shortest-queue (JSQ) scheduling and Random in loss rate, 99th-percentile buffer occupancy, and scalability, while remaining competitive with Top-k at lower hardware cost. PSP provides an effective balance among performance, stability, and overhead for artificial intelligence data centers.
TCP Prague is a congestion control algorithm (CCA) for the low latency, low loss and scalable throughput (L4S) architecture, which targets very low queuing delay by leveraging fine-grained explicit congestion notification (ECN) signals. Most prior evaluations of Prague focus on high-rate scenarios and its TCP-friendliness. Prague can also, however, cope with low-bandwidth bottlenecks or compete with elastic and inelastic flows—leaving only a small fair share per flow. In such cases, Prague reduces its segment size to maintain approximately two packets per virtual round trip time (RTT). The current design derives the segment size from both the pacing rate and RTT, making it sensitive to noisy RTT conditions and potentially degrading performance. We redesign segment sizing to depend solely on the pacing_rate using three schemes—linear, exponential, and logarithmic. We implement these schemes in a modified Prague kernel and evaluate them in our testbed against both the default Prague segment-sizing algorithm and a fixed-segment-size baseline. On throttled links, our schemes reduce retransmissions by more than a factor of 20 and increase feedback granularity. Under both elastic (inter- and intra-CCA) and inelastic competition, they improve throughput stability and fairness.
SprayCast is proposed, a congestion-adaptive native RDMA multicast scheme for dynamic sparse token Dispatch that uses in-band network telemetry feedback to steer replication away from congested multicast branches and range-based negative acknowledgments for localized loss recovery, saving bandwidth and reducing tail latency in dynamic all-to-all communication.
Yingying Zeng, Ning Xu, Xiaobin Tan et al.· Asia-Pacific Workshop on Net...· 0 citations
Data-center congestion control targets high throughput, fair bandwidth allocation, and low latency. Modern transports couple rate computation and packet scheduling into a single feedback loop, converging to near-optimal rates but leaving standing queues that can scale with the number of flows. We argue that separating the two reveals a simpler design point. Given stable feasible rates, the residual queue problem reduces to a timing problem: if every flow's packets arrive at the bottleneck in the correct slot, the link stays busy and the queue stays empty. Clocked ACK-Paced Synchronization CAPS is a lightweight distributed scheduling layer that achieves this by phase-locking each sender's transmissions to ACK-clocked bottleneck slots, with a per-flow correction that compensates for heterogeneous RTTs. We characterize the phase-locked steady state for dumbbell topologies under equal RTT, heterogeneous RTT, and bidirectional traffic, and validate the mechanism on a fat-tree under incast, permutation, and all-to-all traffic. CAPS reduces worst-case queue occupancy by 5-10x across all tested scenarios without throughput loss.
Raphael Zailer, Isaac Keslassy· arXiv.org· 0 citations
A hybrid reinforcement learning (RL) framework that jointly controls queue management and bandwidth allocation in bursty multi-service networks and demonstrates the effectiveness of coordinated learning-based control for stable and QoS-aware operation in bursty networked systems.
T. Khan, Babar Shah, Taimur Karamat et al.· Computing· 0 citations
ProLet is a load balancing scheme that enables proactive probing and reroutes elephant flows at flowlet granularity in lossless RDMA networks and reduces average and tail flow completion time slowdowns by 69% and 79%, respectively, compared to state-of-the-art load balancing schemes.
Hong Wang, Jin-Hao Luo, J. Tan et al.· Asia-Pacific Workshop on Net...· 0 citations
Per-flow measurement of priority-heterogeneous traffic underpins cloud service-level agreement (SLA) enforcement, anomaly detection, and distributed AI training in high-speed networks, yet remains challenging in the fast L1/L2-cache memory regime where high-priority flows are vastly outnumbered. We propose Charon, a priority-aware sketch that replaces the structural separation used by prior methods with stratified admission sampling: a single, online-adaptive, parameter-free rule decides whether each packet is admitted to the sketch. Across multiple real-world traces, Charon achieves more than 2× higher detection accuracy for high-priority flows than the best baseline and up to four orders of magnitude lower average error than state-of-the-art priority-aware sketches, with the gap widening as memory tightens, at high processing throughput. The implementation on the industry-grade Tofino switch further demonstrates low resource utilization.
Weihe Li, Xicheng Li, Dimitrios P. Pezaros et al.· Conference on Applications,...· 0 citations
Related blog posts
MIT News · Artificial Intelligence· news.mit.eduSep 14, 2026
The “HardFlow” algorithm could help generative AI models produce high-quality outputs that obey strict requirements when “pretty close” doesn’t cut it.
AI may appear weightless, but every model depends on physical infrastructure. To understand responsible AI, we need to look beyond algorithms and consider the entire lifecycle of the hardware behind them. The post Responsible AI Must Consider Its Afterlife appeared first on GPT-Lab.
