Synchronization-induced congestion is a critical performance bottleneck in modern distributed machine learning (ML) training, where simultaneous gradient exchanges create bursty traffic patterns. Existing solutions, both reactive and proactive, struggle to balance throughput and latency in the presence of synchronized flows. We propose DeSync, a proactive traffic shaping scheme that introduces structured random delay to de-synchronize communication rounds. Evaluations with DCQCN, HPCC, DCTCP, and TIMELY demonstrate that DeSync significantly improves FCT, job completion times, and congestion metrics, enhancing existing CC mechanisms without specialized hardware.

Communication overhead is a major bottleneck in distributed AI training, particularly in multi-tenant environments, limiting GPU utilization. Existing job-level scheduling methods fail to address the varying urgency of individual communication operations. We propose Flux, a novel fine-grained scheduler that prioritizes communication operations based on their Urgency Score and job intensity. Our evaluation shows Flux improves GPU utilization by up to 10 % compared to state-of-the-art job-level algorithms. This demonstrates the significant advantage of fine-grained communication scheduling in multitenant AI clusters.

While Multi-CQF enables deterministic networking over wide-area networks(WANs) by decoupling transmission and reception, it introduces significant queuing delays. We propose Effective Phase Alignment (EPA), which mitigates queuing delay by adjusting transmission offsets to align the effective phase, defined as the phase difference between the sending and receiving windows. EPA lowers the upper bound of average queuing delay from 2T to 1.5T, and achieves T under perfect alignment.

Integrating Frame Replication and Elimination for Reliability (FRER) with Cyclic Queuing and Forwarding (CQF) in Time-Sensitive Networks (TSN) encounters redundancy failures and resource reservation inefficiencies due to length disparities across redundant paths. To address these challenges, we propose ReCQF, a Reliability-Enhanced CQF scheduling framework built on Multi-Instance CQF. ReCQF adaptively assigns redundant flows to multiple CQF queue pairs with specific cycles, effectively aligning transmission delays across redundant paths to ensure low delay and inter-path delay differences while significantly reducing resource reservations.

Time-sensitive networking (TSN), essential in industrial networks for its promise of reliable and deterministic data transmission, faces deployment challenges due to the limitations of existing time-aware shaper (TAS)-based scheduling algorithms. Specifically, the size of the generated gate control lists (GCLs) is usually too large to be deployed in actual devices. To bridge the gap between theory and practice, we propose FlexTAS, a flexible and practical solution for TSN. The key insight behind FlexTAS is that relaxing gating does not introduce uncertainty, as long as nonoverlap reserved time slots are guaranteed. FlexTAS is comprised of two main components: first, a novel gating model deviates from the conventional TAS model by incorporating selective relaxation of gating at certain nodes; and second, a deep reinforcement learning-based engine to rapidly generate valid schedules. We build a real testbed and validate the effectiveness of our proposed solution. Our evaluation demonstrates that FlexTAS effectively controls the number of gate entries within the GCL capacity of devices, while simultaneously meeting the Quality of Service(QoS) requirements of time-triggered streams. It significantly reduces the number of GCL entries by 60% to 80%, and facilitates deployment in heterogeneous networks, thus offering a practical solution for TSN.

Industrial networks require deterministic and reliable communication, which can be achieved by Time-Sensitive Networking (TSN), a set of standards that enable precise timing and synchronization of data transmission. However, TSN is susceptible to unexpected delays caused by device malfunction, interference or cyber attacks, which can have a domino effect and disrupt multiple data flows. To address this challenge, we propose RobustTSN, a framework that protects TSN against the domino effect of delayed frames and tolerates harmless accident frames using Per-Stream Filtering and Policing (PSFP) mechanism. We develop algorithms to calculate ingress filtering schedules based on local-safe delay and global-safe interval concepts, which decide whether to accept or discard out-of-schedule frames. We use a finite state machine to model the interaction between frames and evaluate frame safety. We build a software-defined networking based system to dynamically monitor network states and reconfigure device filtering after out-of-schedule transmission occurs. We conduct experiments on practical scenario topologies and large groups of random flows to demonstrate the effectiveness and efficiency of our framework.

In the domain of Time-Sensitive Networking (TSN), the quest for ultra-reliable low-latency communication is paramount. Current scheduling strategies, which hinge on strict isolation to ensure low latency and jitter, confront the challenges of high overhead in worst-case latency evaluation and consequent limitations in network flow capacity. This paper introduces an innovative framework that transcends traditional isolation constraints, thereby expanding the solution space and augmenting network schedulability. At the heart of this framework lies a novel latency jitter analysis method that assesses the viability of non-isolation scenarios with constant time complexity. This method underpins a heuristic scheduling algorithm that not only boasts the smallest time complexity among existing heuristics but also significantly increases the number of scheduled flows. Complementing this, we integrate a discrete time reference approach to hasten time-intensive scheduling operations, achieving an optimal balance between schedulability and runtime efficiency. The framework further incorporates a workload-shifting technique to enhance online scheduling responsiveness. It adeptly manages the variability in scheduling times caused by disharmonious flow periods, further bolstering the framework’s robustness. Experimental validations demonstrate that our framework can increase the scheduled flows up to 269%. It reduces scheduling runtime by up to 98.44% for medium-scale networks while maintaining a flat runtime growth curve, ensuring predictable performance in online scheduling scenarios.

5G and Industrial Internet are bringing a variety of applications with on-time and reliable demands. Cyclic queuing and forwarding (CQF), a well-known mechanism defined by IEEE 802.1 Qch in Time Sensitive Network (TSN), achieves deterministic end-to-end latency and jitter without complex gating calculations. However, most of the current work ignores the prevalence of hybrid networks with different link rates, resulting in low bandwidth utilization and high latency for single-cycle CQF. In this paper, we propose a multi-cycle CQF to address the transmission in multi-link-rate networking, reducing deterministic end-to-end latency and improving link bandwidth utilization. In addition, we formulate the scheduling constraints, being of guiding significance for designing the transmission of multi-link-rate networks, and we design an online scheduling algorithm based on it. We compare the proposed scheme with the single-cycle CQF online scheduling algorithm in hierarchical multi-link-rate networking scenarios, and the evaluation shows that our algorithm achieves better end-to-end ultra-low latency (38.9% reduction) with a smaller schedulability gap compared with single-cycle CQF.

Time-Sensitive Networking (TSN) is an emerging network architecture that provides bounded latency and reliable network services for time-sensitive applications. Since time-triggered flows in TSN are typically periodic, a concept of network cycle is widely used in both standards and academic researches. However, although network cycle has gained popularity, its rationale has not yet been analyzed systematically.In this paper, we mathematically evaluate the performance of several flow scheduling algorithms in terms of flow schedulability with and without employing network cycle. We observe that only when the network cycle is set to a proper value can the performance of flow scheduling be significantly improved. To better evaluate the scheduling effect, a novel assessment metric and a goal-based optimization algorithm are introduced. Our experiment results show that the network cycle-based algorithm can achieve a considerable improvement (40% - 170% improvement in the number of scheduled flows) compared to the ones with network cycle disabled.