Huanle Xu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Huanle Xu has a total of 84 co-authors. Here are the top co-authors by number of shared publications:
- Cheng-Zhong Xu
State Key Laboratory of IoTSC, University of Macau, Macau, Macao
14 shared publications - Wing Cheong Lau
The Chinese University of Hong Kong
13 shared publications - Zizhao Mo
University of Macau, Macau, China
8 shared publications - Kejiang Ye
Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences
7 shared publications - Yang Liu
Shanghai University
5 shared publications - Shutian Luo
University of Virginia, Charlottesville, USA
4 shared publications
Latest Publications
- Open Access
- 18 November 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 24 (1) , 1-26
AdaFuse: Adaptive Function Fusion for Dynamic Serverless Workflows
Serverless workflows are commonly modeled as directed acyclic graphs (DAGs) composed of fine-grained functions. While this design promotes modularity and programming flexibility, it also incurs frequent inter-function communication. To alleviate this performance overhead, function fusion has been explored to co-locate and execute communicating functions within a shared container instance. However, existing approaches fail to account for the time-varying tradeoff between communication latency reduction and resource overprovisioning arising from path uncertainty in dynamic serverless workflows, making it difficult to sustain high resource efficiency over time. In this article, we propose a new adaptive function fusion strategy that dynamically reconstructs serverless workflows by adjusting communication schemes between functions in response to changing workflow invocations. Realizing this strategy, however, requires jointly optimizing function fusion and per-instance configuration, a problem with a combinatorial search space, while managing the substantial overhead of switching fusion schemes at runtime. To address these challenges, we design AdaFuse, a novel system for dynamic serverless workflows. AdaFuse jointly optimizes function fusion and instance reconfiguration through an efficient beam-search–based algorithm to optimize the end-to-end resource allocation. With the optimized fusion scheme, AdaFuse employs a fluid migration mechanism that incrementally transitions the workflow deployment while seamlessly redirecting incoming workloads. Evaluations across diverse benchmarks show that AdaFuse reduces resource allocation by up to 49.4% compared to baselines, without compromising SLO attainment.
- 18 November 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 24 (1) , 1-26
Open AccessAdaFuse: Adaptive Function Fusion for Dynamic Serverless Workflows
Serverless workflows are commonly modeled as directed acyclic graphs (DAGs) composed of fine-grained functions. While this design promotes modularity and programming flexibility, it also incurs frequent inter-function communication. To alleviate this performance overhead, function fusion has been explored to co-locate and execute communicating functions within a shared container instance. However, existing approaches fail to account for the time-varying tradeoff between communication latency reduction and resource overprovisioning arising from path uncertainty in dynamic serverless workflows, making it difficult to sustain high resource efficiency over time. In this article, we propose a new adaptive function fusion strategy that dynamically reconstructs serverless workflows by adjusting communication schemes between functions in response to changing workflow invocations. Realizing this strategy, however, requires jointly optimizing function fusion and per-instance configuration, a problem with a combinatorial search space, while managing the substantial overhead of switching fusion schemes at runtime. To address these challenges, we design AdaFuse, a novel system for dynamic serverless workflows. AdaFuse jointly optimizes function fusion and instance reconfiguration through an efficient beam-search–based algorithm to optimize the end-to-end resource allocation. With the optimized fusion scheme, AdaFuse employs a fluid migration mechanism that incrementally transitions the workflow deployment while seamlessly redirecting incoming workloads. Evaluations across diverse benchmarks show that AdaFuse reduces resource allocation by up to 49.4% compared to baselines, without compromising SLO attainment.
- Open Access
- 14 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 1499-1513
Axon: Asynchronous Global KV-Aware Scheduling Across Distributed Instances for Agentic Workloads
- 14 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 1499-1513
Open AccessAxon: Asynchronous Global KV-Aware Scheduling Across Distributed Instances for Agentic Workloads
- Open Access
- 5 November 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
PLUGSCHED+: Efficient and General Live Update and Recovery for Linux Scheduler
The scheduler is a critical component of the operating system (OS) and is tightly coupled with Linux. Production-level clouds often host various workloads, and these workloads require different schedulers to achieve high performance. Furthermore, the scheduler may encounter faults during runtime, leading to the kernel panic. Thus the capability of updating and recovering the scheduler lively without rebooting the OS is crucial for the production environments. However, existing techniques only apply for the fine-grained function-level updates, or require extra constraints such as virtualization when performing update and recovery. We therefore propose Plugsched+ to enable scheduler live update and recovery with two key novelties. First, with the idea of modularization, Plugsched+ decouples the scheduler from the Linux kernel to be an independent module. Second, Plugsched+ uses the data rebuild technique to migrate the state from the old scheduler to the new one during live update, or clear anomalous data during live recovery. This scheme can be directly applied to the Linux kernel scheduler in production environments without replacing the kernel. Experimental and production results show that Plugsched+ can effectively update and recover the scheduler lively, and the downtime is no more than tens of milliseconds.
- 5 November 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
Open AccessPLUGSCHED+: Efficient and General Live Update and Recovery for Linux Scheduler
The scheduler is a critical component of the operating system (OS) and is tightly coupled with Linux. Production-level clouds often host various workloads, and these workloads require different schedulers to achieve high performance. Furthermore, the scheduler may encounter faults during runtime, leading to the kernel panic. Thus the capability of updating and recovering the scheduler lively without rebooting the OS is crucial for the production environments. However, existing techniques only apply for the fine-grained function-level updates, or require extra constraints such as virtualization when performing update and recovery. We therefore propose Plugsched+ to enable scheduler live update and recovery with two key novelties. First, with the idea of modularization, Plugsched+ decouples the scheduler from the Linux kernel to be an independent module. Second, Plugsched+ uses the data rebuild technique to migrate the state from the old scheduler to the new one during live update, or clear anomalous data during live recovery. This scheme can be directly applied to the Linux kernel scheduler in production environments without replacing the kernel. Experimental and production results show that Plugsched+ can effectively update and recover the scheduler lively, and the downtime is no more than tens of milliseconds.
- 16 October 2026
- Published by Wiley in Software: Practice and Experience
Cloud‐Native and Distributed Systems for Efficient and Scalable Large Language Models—A Research Agenda
Objective: Large Language Models (LLMs) are rapidly transforming artificial intelligence applications, but their unprecedented computational, memory, communication, and data‐management demands pose significant challenges to existing computing infrastructures. Traditional cloud and distributed systems were not designed for the scale, dynamic execution patterns, heterogeneous hardware requirements, and service‐level demands of modern LLM workloads. This paper aims to examine how cloud‐native and distributed systems can evolve to support efficient, scalable, reliable, and sustainable LLM training and inference. Methods: We conduct a structured analysis of representative research advances and practical developments in distributed LLM training, inference serving, cloud‐native orchestration, resource management, heterogeneous computing, and edge‐cloud intelligence. We organize the analysis around major system challenges, including computation, distributed execution and system software, resource management and serving, security and privacy, data management, and standardization. We further examine representative resource‐management strategies and emerging paradigms, including serverless inference, federated and decentralized learning, trustworthy distributed AI, and heterogeneous computing with emerging accelerators. Results: Our analysis reveals that supporting LLMs at scale requires a transition from general‐purpose cloud infrastructure toward LLM‐aware, cross‐layer, and adaptive systems. Existing approaches have improved distributed execution, elastic resource management, multi‐tenant serving, and hardware utilization, but important gaps remain in cross‐layer coordination, workload‐aware orchestration, data and context management, interoperability, security, and system‐level optimization. Based on these gaps, we develop a research agenda covering future directions in sparse and heterogeneous computation, LLM‐enhanced system software, prompt and context management, autonomous cloud management, decentralized intelligence, edge‐cloud collaboration, personalized LLM services, and emerging computing paradigms. Conclusion: Cloud‐native and distributed systems will play a central role in enabling the next generation of LLM services. Future infrastructures should evolve toward more intelligent, elastic, interoperable, secure, and heterogeneous platforms that jointly optimize computation, memory, communication, data, and service objectives. Continued advances in system design, standardization, and cross‐disciplinary collaboration will be essential for realizing scalable and sustainable LLM infrastructures.
- 16 October 2026
- Published by Wiley in Software: Practice and Experience
Cloud‐Native and Distributed Systems for Efficient and Scalable Large Language Models—A Research Agenda
Objective: Large Language Models (LLMs) are rapidly transforming artificial intelligence applications, but their unprecedented computational, memory, communication, and data‐management demands pose significant challenges to existing computing infrastructures. Traditional cloud and distributed systems were not designed for the scale, dynamic execution patterns, heterogeneous hardware requirements, and service‐level demands of modern LLM workloads. This paper aims to examine how cloud‐native and distributed systems can evolve to support efficient, scalable, reliable, and sustainable LLM training and inference. Methods: We conduct a structured analysis of representative research advances and practical developments in distributed LLM training, inference serving, cloud‐native orchestration, resource management, heterogeneous computing, and edge‐cloud intelligence. We organize the analysis around major system challenges, including computation, distributed execution and system software, resource management and serving, security and privacy, data management, and standardization. We further examine representative resource‐management strategies and emerging paradigms, including serverless inference, federated and decentralized learning, trustworthy distributed AI, and heterogeneous computing with emerging accelerators. Results: Our analysis reveals that supporting LLMs at scale requires a transition from general‐purpose cloud infrastructure toward LLM‐aware, cross‐layer, and adaptive systems. Existing approaches have improved distributed execution, elastic resource management, multi‐tenant serving, and hardware utilization, but important gaps remain in cross‐layer coordination, workload‐aware orchestration, data and context management, interoperability, security, and system‐level optimization. Based on these gaps, we develop a research agenda covering future directions in sparse and heterogeneous computation, LLM‐enhanced system software, prompt and context management, autonomous cloud management, decentralized intelligence, edge‐cloud collaboration, personalized LLM services, and emerging computing paradigms. Conclusion: Cloud‐native and distributed systems will play a central role in enabling the next generation of LLM services. Future infrastructures should evolve toward more intelligent, elastic, interoperable, secure, and heterogeneous platforms that jointly optimize computation, memory, communication, data, and service objectives. Continued advances in system design, standardization, and cross‐disciplinary collaboration will be essential for realizing scalable and sustainable LLM infrastructures.
- 6 October 2026
- Published by Springer Nature
- p. 147-163
AI-Lite: Adaptive Data Center Cooling Optimization Based on Machine Learning
- 6 October 2026
- Published by Springer Nature
- p. 147-163
AI-Lite: Adaptive Data Center Cooling Optimization Based on Machine Learning
- Open Access
- 13 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 292-304
Cremes: Cost-Efficient and Reliable Microservice Execution on Spot Instances
- 13 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 292-304
Open AccessCremes: Cost-Efficient and Reliable Microservice Execution on Spot Instances
- Open Access
- 18 May 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Management of Data
- Vol. 4 (3) , 1-26
Serving Hybrid LLM Loads with SLO Guarantees Using CPU-GPU Attention Piggybacking
Nowadays, service providers often deploy multiple types of LLM services within shared clusters. While the service co-location improves resource utilization, it introduces significant interference risks for latency-sensitive (LS) services -- which have strict SLO requirements for inference latency -- and severely constrains the service capacity of best-effort (BE) services due to limited available memory. To address interference, existing systems typically rely on reserving headroom to constrain BE resource usage. However, this approach's coarse granularity compromises the SLO compliance of the latency-sensitive service and unnecessarily restricts the generation potential of the best-effort service. In this paper, we propose OmniServe, a novel LLM serving system that efficiently harnesses both CPU and GPU resources to mitigate interference and improve throughput. Central to OmniServe is the Attention Piggybacking mechanism, which effectively offloads the Attention computation of BE services to CPUs on the fly. This mechanism also facilitates asynchronous communication between CPU and GPU streams, preventing GPUs from being blocked while aggregating Attention results. Additionally, OmniServe incorporates a dynamic batching control policy to adapt to fluctuating request arrivals, facilitating Dense module computation using layer-wise batching. Experimental results show that OmniServe improves the SLO attainment rate for LS services by up to 1.48× while enhancing BE serving throughput by up to 9.85× compared to state-of-the-art systems.
- 18 May 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Management of Data
- Vol. 4 (3) , 1-26
Open AccessServing Hybrid LLM Loads with SLO Guarantees Using CPU-GPU Attention Piggybacking
Nowadays, service providers often deploy multiple types of LLM services within shared clusters. While the service co-location improves resource utilization, it introduces significant interference risks for latency-sensitive (LS) services -- which have strict SLO requirements for inference latency -- and severely constrains the service capacity of best-effort (BE) services due to limited available memory. To address interference, existing systems typically rely on reserving headroom to constrain BE resource usage. However, this approach's coarse granularity compromises the SLO compliance of the latency-sensitive service and unnecessarily restricts the generation potential of the best-effort service. In this paper, we propose OmniServe, a novel LLM serving system that efficiently harnesses both CPU and GPU resources to mitigate interference and improve throughput. Central to OmniServe is the Attention Piggybacking mechanism, which effectively offloads the Attention computation of BE services to CPUs on the fly. This mechanism also facilitates asynchronous communication between CPU and GPU streams, preventing GPUs from being blocked while aggregating Attention results. Additionally, OmniServe incorporates a dynamic batching control policy to adapt to fluctuating request arrivals, facilitating Dense module computation using layer-wise batching. Experimental results show that OmniServe improves the SLO attainment rate for LS services by up to 1.48× while enhancing BE serving throughput by up to 9.85× compared to state-of-the-art systems.
- Open Access
- 26 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 56-71
High Throughput and Low Latency LLM Serving via Adaptive KV Caching
- 26 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 56-71
Open AccessHigh Throughput and Low Latency LLM Serving via Adaptive KV Caching
- Open Access
- 19 November 2025
- Published by Association for Computing Machinery (ACM)
- p. 521-534
FedDance: Efficient Participant Selection for Federated Learning in Highly Dynamic Environments
- 19 November 2025
- Published by Association for Computing Machinery (ACM)
- p. 521-534
Open AccessFedDance: Efficient Participant Selection for Federated Learning in Highly Dynamic Environments