To satisfy the unprecedented Quality of Experience (QoE) and stringent latency Service Level Agreements (SLAs) of emerging interactive applications, modern Content Delivery Networks (CDNs) are deploying massively decentralized edge nodes. However, this paradigm shift poses a significant challenge: optimal traffic scheduling fundamentally depends on acquiring real-time, full-coverage end-to-end path latency data to prevent SLA violations. Current measurement methods cannot scale to monitor every possible user-to-node path, and traditional prediction approaches (e.g., relying on additional segmented measurements or low-rank matrix decomposition) fail to achieve satisfactory accuracy on the resulting extremely sparse datasets. In this work, we present ADePT (Application Delay PredicTion), a novel data-driven causal inference framework that provides comprehensive and precise latency predictions without requiring additional measurements. By explicitly decoupling user-side temporal variations (e.g., last-mile congestion) and node-side spatial variations (e.g., core propagation delays), ADePT successfully extracts high-dimensional latent embeddings from limited measurement data to infer the end-to-end path latency for any potential scheduling decision. Evaluated on a massive real-world dataset from a leading edge CDN, ADePT reduces prediction errors by 19.6% and achieves a median absolute error of 4.3 ms. Consequently, integrating ADePT's accurate predictions into CDN traffic scheduling significantly improves scheduling decisions, increasing the ratio of traffic meeting strict applications' latency requirements by 1.66 & times;.

Low-latency cloud rendering enables mobile users to experience high-quality, real-time 3D graphics but achieving low Motion-to-Photon (MTP) latency while maintaining smooth playback is a significant challenge. Our real-world measurement study identifies Receive-to-Composition (R2C) latency, caused by ineffective jitter buffer management, as the primary factor contributing to increased MTP latency. To address this, we introduce JitBright, a client-side optimization strategy that dynamically reduces MTP latency through adaptive jitter buffer management. By adjusting buffer levels based on smoothing playback probability and implementing proactive keyframe requests to mitigate frame dependency, JitBright minimizes both active and passive waiting times.

Real-time broadband communication (RTBC) scenarios, such as cloud virtual reality and 8K live streaming, further raise the criteria of the performance triangle, requiring video bitrates exceeding 30 Mbps, tail delay below 50 ms, and fairness guarantees for multi-user concurrent access. Based on our testing and analysis, existing RTBC-oriented rate control solutions, including end-to-end algorithms and network-assisted algorithms, fail to simultaneously satisfy all performance metrics. The native dynamic delay and physical-layer resource allocation strategy inherent to the 5G radio access network (RAN) are the key reasons. These solutions lack adaptation to the 5G architecture, leading to reduced decision performance. This paper proposes Choir, an innovative collaborative solution mainly deployed on 5G base stations that deeply integrates 5G radio characteristics and video streaming traffic patterns to guide efficient sender-side rate control. Extensive simulation and testbed evaluations demonstrate Choir's significant performance in achieving high average bitrate, low tail delay, and inter-flow fairness across different 5G network scenarios.

Deep Learning Training (DLT) jobs have become the major workload in cloud GPU clusters. As an increasing variety of distinctly different DLT jobs are trained simultaneously on clusters and these jobs adopt different parallelism strategies, the challenge lies in how to schedule these jobs to efficiently utilize cluster resources considering their communication patterns. In this paper, we introduce Jaca, a job scheduler aiming to reduce Job Completion Time (JCT) by selecting the job with the highest affinity-a metric we define to reflect how efficiently a job can run on the current cluster-based on jobs' communication patterns and current cluster loads. To efficiently calculate the affinity within a reasonable time, we group the machines of the cluster, sample and prune placement orders to reduce the exploded placements plan space. This approach comprehensively considers the impact on the cluster, ultimately scheduling jobs along with their placement. Evaluation on a 12-GPU cluster demonstrates that Jaca improves the average JCT by 1.40 & times;similar to 2.82 & times;. The large-scale trace-based simulation further demonstrates that Jaca improves average JCT by up to 1.45 & times;similar to 2.40 & times; compared with the traditional job schedulers like RL-Hybrid, Yarn, Tiresias, and Gandiva.

The rise of large language models (LLMs) has led to intense communication contention in multi-job GPU training clusters. Existing communication schedulers typically adopt a resource-centric perspective, prioritizing flows based on static metrics, such as GPU intensity. However, this approach ignores a fundamental complexity: the entanglement of intra-job execution dependency and inter-job network contention. In this paper, we present Moirai, a communication scheduler adopting a dependency-impact-based scheduling. Moirai explicitly models the entanglement and is built on two core innovations to address the system’s circular causality and vast search space: (1) a Dependency Resolution Engine that uses an analytical fixed-point iteration model to decouple the entangled dependency loops and precisely quantify the global impact of each flow; and (2) an Impact-Guided Strategy Explorer that navigates the exploded strategy space using heuristic optimization signals to identify flow-level strategies. Evaluations in a high-fidelity, ns-3-based simulator show that Moirai improves average Job Completion Time (JCT) by 1.38× – 1.52× and boosts GPU utilization by 13.8

Large-scale edge Content Delivery Networks (CDNs) provide low-latency content access services and suffer from high bandwidth costs. While previous studies have sought to optimize bandwidth costs under percentile billing, the efficacy is compromised due to the pervasive uncertainty inherent in practical systems, including traffic demand dynamics, performance-constrained scheduling bias, and systemic scheduling deviations. Such uncertainties can result in large gaps among optimal, expected, and actual utilization of massive vulnerable and heterogeneous edge nodes. To address these uncertainties, we propose Oceanus, a cost-effective traffic scheduling system for large-scale edge CDN systems. Oceanus decouples the bandwidth planning problem and performs on multiple timescales. In addition, Oceanus coordinates bandwidth planning with flow scheduling through the bidirectional feedback scheme. Oceanus further utilizes nodes with minimal marginal cost to reduce additional bandwidth cost. Extensive experiments in a trace-driven testbed and real-world deployment confirm the effectiveness of Oceanus. Compared to the state-of-the-art scheduling method, Oceanus achieves 79.4% (vs. 51.5%) of optimum cost reduction and reduces 21.4% (vs. 8.1%) bandwidth costs.

Mobile e-commerce platforms increasingly integrate cloud rendering to deliver immersive 3D shopping experiences, where users interact with the rendered scenes through the network. Our large-scale online measurements reveal that users' Quality of Experience (QoE) preferences dynamically evolve with user motions in cloud rendering sessions. However, latency spikes occur more frequently during peak periods of user engagement, resulting in early session abandonment. To address this issue, we propose MARC, a motion-aware rate control framework that aligns bitrate decisions with user QoE preferences in real-time. MARC sets dynamic QoE objectives based on real-world user engagement behavior, captures the different latency and quality requirements for motion and non-motion frames, and employs stochastic optimization to maximize QoE. Extensive deployment of over 1 million user sessions demonstrates that MARC reduces session freeze rates by 71% and increases user interaction time by 20%, significantly improving user engagement for e-commerce cloud rendering.

The live streaming landscape has shifted to a crowd-sourced paradigm, resulting in highly volatile and geographically diverse viewer demand. To handle growing traffic, Content Delivery Networks (CDNs) increasingly rely on a mix of dedicated infrastructure and lower-cost, heterogeneous edge resources. Our analysis of production data reveals two emerging characteristics in modern live delivery: dynamic regional supply-demand imbalance and per-stream heterogeneity in popularity and geography. However, existing request mapping solutions fall short in this new landscape, as they assume stable regional capacity and overlook stream-level heterogeneity. This paper proposes LiveMap , a cost-efficient and latency-aware request mapping system for live CDNs. LiveMap performs online bandwidth provisioning via dual-level coordination to resolve regional supply-demand imbalances and reduce bandwidth costs. Further, LiveMap introduces an adaptive stream mapping strategy that dynamically forms per-stream delivery groups based on real-time popularity and system load. Deployed in Bilibili CDN serving crowdsourced live streaming over a year, LiveMap reduces bandwidth costs by 42.18% and access latency by 20.26%, outperforming the state-of-the-art solutions.

Real-time video (RTV) systems place high demands on ultra low-latency (i.e., less than 100 ms). However, our large-scale measurements reveal that a significant portion of users still experience high video frame latency due to bandwidth jitters. Existing solutions attempt to mitigate this issue by lowering the sender's future video frame encoding bitrate. Nevertheless, as shown in our controlled experiments, they fail to drain existing packets queued on the bottleneck node (i.e., the 5G base station and Wi-Fi access point), still suffering from high tail latency as bandwidth decreases. In this paper, we propose Co-RTV, a collaborative RTV system that achieves predictable latency control. Specifically, Co-RTV enables endpoint-network collaboration between the bottleneck node and the sender. The collaboration speeds up the release of packets queued at the bottleneck node and facilitates accurate latency control at the RTV sender through scalable QoE-driven flow control. Extensive experiments in emulated networks and on a 5G testbed demonstrate the superior performance of Co-RTV, with tail latency reductions of 69.1% and 70.5%, respectively.

Large language models (LLMs) have recently demonstrated notable performance, particularly in addressing the challenge of extensive data requirements when training traditional forecasting models. However, these methods encounter significant challenges when applied to high-dimensional and domain-specific datasets. These challenges primarily arise from inability to effectively model inter-variable dependencies and capture variable-specific characteristics, leading to suboptimal performance in complex forecasting scenarios. To address these limitations, we propose ADTime, an adaptive LLM-based approach for multivariate time series forecasting. ADTime employs advanced preprocessing techniques to identify latent relationships among key variables and temporal features. Additionally, it integrates temporal alignment mechanisms and prompt-based strategies to enhance the semantic understanding of forecasting tasks by LLMs. Experimental results show that ADTime outperforms state-of-the-art methods, reducing MSE by 9.5% and MAE by 6.1% on public datasets, and by 17.1% and 13.5% on domain-specific datasets. Furthermore, zero-shot experiments on real-world refinery datasets demonstrate that ADTime exhibits stronger generalization capabilities across various transfer scenarios. These findings highlight the potential of ADTime in advancing complex, domain-specific time series forecasting tasks.

Emerging mobile short video services pose different yet stringent performance requirements compared to traditional long video services. Content providers (CPs) aspire to a better user-perceived Quality of Experience (QoE) at the application layer, which is imperceptible to the Content Delivery Network (CDN), which monitors Quality of Service (QoS) at the transport layer. The mismatch between QoS and QoE leads to a complex and diverse mapping correlation between the two metrics. In this paper, we illustrate the QoS-QoE mapping correlation in mobile short video services. Although data-driven QoE prediction models can achieve the desired accuracy, complex scenario features are proven to be necessary, and the prediction model still lacks interpretability. Deeper quantitative analysis shows that the correlation becomes complex and diverse when resources are insufficient. The clustering-based prediction framework can successfully summarize scenario features and perform QoE prediction based on QoS metrics alone. Furthermore, we propose predictive QoE-based CDN scheduling. Experiments show that compared to scheduling with QoS metrics, QoE-aware scheduling achieves an average QoE improvement of 9.9

In this paper, we present TECC, a system based on collaborative transmission control that mitigates the mismatch of sending behavior between the inner and outer connections to achieve efficient QUIC tunneling. In TECC, a feedback framework is implemented to enable end hosts to collect more precise network information that is sensed on the tunnel server, which assists the inner end-to-end connection to achieve better congestion control and loss recovery. Extensive experiments in emulated networks and real-world large-scale A/B tests demonstrate the efficiency of TECC. Specifically, compared with the state-of-the-art QUIC tunneling solution, TECC significantly reduces flow completion time. In emulated networks, TECC decreases flow completion time by 30% on average and 53% at the 99th percentile. TECC also gains a reduction in RPC (Remote Procedure Call) request completion time of 3.9% on average and 13.3% at the 99th percentile in large-scale A/B tests.

Distributed machine learning (DML) has recently experienced widespread application. A major performance bottleneck is the costly communication for gradients synchronization. Recently, researchers have explored the use of programmable switches for in-network synchronous aggregation of gradients to mitigate the communication overhead. Nevertheless, the performance of in-network synchronous aggregation is significantly impacted by the stragglers. Unfortunately, the schedulers in existing DML systems are no longer effective in dealing with stragglers because of the ignorance of the aggregation progress that is offloaded from the parameter servers to the programmable switches. To address this gap, this paper presents VAKY, an adaptive scheduler specifically designed for in-network aggregation. At the heart of VAKY is the variable K-block sync method, where the aggregators stop waiting for updates from more workers once having received updates from the fastest K workers for each block of gradients. We propose an efficient solution that can dynamically choose the optimal values of K during the training process, in order to minimize the expected training completion time. We have integrated VAKY into PyTorch, and our experiments show that compared to the state-of-the-art in-network aggregation systems, VAKY improves the aggregation throughput by up to 40% and reduces the training time by 25%.

Virtual Reality (VR) applications that require extremely low delay and high image quality are widely used in online games and other 5G scenarios, becoming a hot research field in recent years. However, the limited bandwidth in 5G edge networks fails to meet the peak rate requirements for multiple VR flows. MPTCP is suitable for 5G edge networks, supporting the simultaneous use of multiple networks on mobile devices. Nevertheless, accurately scheduling VR data blocks to different subflows to satisfy their low delay requirements is challenging due to their microburst characteristic. In this paper, we propose a novel MPTCP scheduler for cloud VR applications in 5G edge networks, called the Cross-layer information-based one-way delay Predictive Scheduler (CPS). CPS accurately predicts one-way delay by incorporating cross-layer information from both the application and edge wireless sides, and adaptively schedules VR data blocks to the optimal subflow. Experimental results show that CPS outperforms existing strategies, supporting 125% more users for VR applications in the typical scenario. CPS successfully meets the quality of experience needs of more users, providing a promising solution for large-scale deployment of cloud VR services in 5G edge networks.

In real-time communication (RTC) systems, accurate bandwidth prediction is crucial for encoding and transmission strategies to optimize users' quality of experience (QoE) in various network environments. In this paper, we propose an offline reinforcement learning (RL) method to predict bandwidth for RTC video streaming. We use a representative algorithm, named Implicit Q-Learning (IQL), to train the model. To improve the performance, we carefully preprocess the given dataset and redesign the neural network structure and the reward function. Ablation studies are performed to verify our design choices. Furthermore, compared to a baseline method and six behavior policies, our method reduces the mean squared error (MSE) by 18%-22%, demonstrating high prediction accuracy. Our proposed method won the first prize in ACM MMSys 2024 Grand Challenge on Offline Reinforcement Learning for Bandwidth Estimation in Real Time Communications. The source code is available at https://github.com/n13eho/Schaferct.

Access networks, even with advanced 5G technology, often face bottlenecks when supporting concurrent real-time Ultra High Definition (UHD) video streams with high bandwidth and low latency (e.g., under 10 ms of one-way delay) requirements. Traditionally, end systems employ a combination of flow and congestion control mechanisms to control the sending rate to avoid overwhelming the receiver and the network. However, such control efforts induce prolonged tail delays, thereby sharply reducing the number of UHD video streams meeting delivery deadlines, and sometimes even zero. These outcomes are largely due to the inaccurate network status estimation associated with the control mechanisms. To address this challenge, we propose CFC, a deadline-oriented flow control mechanism that employs cross-layer status estimation to maximize user satisfaction with deadlines. CFC accurately assesses cross-layer information, including flow status and 5G access network status at minimal expense, thus ensuring the deadlines through effective concurrent flow control. Our experiments, conducted in both simulation and testbed settings, demonstrate significant improvements in delay and load-balancing for both reliable and unreliable transmissions.