Wenfei Wu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Research Interests
Wenfei Wu has a total of 181 co-authors. Here are the top co-authors by number of shared publications:
- Yuchen Xu
Peking University, Beijing, China
8 shared publications - Jianglong Nie
Peking University, Beijing, China
5 shared publications - Yitao Yuan
Peking University, Beijing, China
4 shared publications - Jiangyuan Chen
School of Computer Science and Technology, University of Science and Technology of China, Hefei, China
3 shared publications - Shaoke Fang
Peking University, Beijing, China
3 shared publications - Xiangrui Yang
College of Computer Science and Technology, National University of Defense and Technology, Changsha, China
3 shared publications
Latest Publications
- Open Access
- 14 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 1180-1195
AgentMesh: Toward Redundancy- and Barrier-Free Dataflow in Multi-Agent Systems
- 14 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 1180-1195
Open AccessAgentMesh: Toward Redundancy- and Barrier-Free Dataflow in Multi-Agent Systems
- Open Access
- 28 September 2026
- Published by Association for Computing Machinery (ACM)
- p. 1091-1106
M2K: Making the Model-Kernel Interface Explicit for Reliable CUDA Kernel Verification
Large language model (LLM) inference systems rely on CUDA kernels for core GPU computations, yet the interface between models and kernels is implicit and poorly specified. Models and kernels evolve independently and often make incompatible assumptions about tensor shapes and input sizes, leading to subtle memory bugs in CUDA kernels. These bugs can crash inference services, corrupt model weights, or be exploited by remote adversaries. Existing techniques either incur prohibitive runtime overhead, require specialized hardware, or fail to handle dynamic tensor shapes and variable kernel launch configurations, leaving the CUDA memory bugs largely unaddressed. This paper presents M2K, a fully automated framework that makes the model-kernel interface explicit and leverages it to detect memory bugs in CUDA kernels used in LLM inference systems. M2K consists of two components. HFProbe traces model execution without GPU hardware, classifies kernel arguments into model-fixed and user-variable, and emits symbolic constraints that capture the interface. cuK-LEE then performs symbolic execution on CUDA kernels to pinpoint memory bugs under the interface constraints, modeling tensors as disjoint memory regions and treating thread identifiers symbolically to scale to thousands of threads. In the evaluation, M2K discovers 181 previously unknown bugs in real LLM inference systems, while producing only nine false positives, demonstrating its effectiveness.
- 28 September 2026
- Published by Association for Computing Machinery (ACM)
- p. 1091-1106
Open AccessM2K: Making the Model-Kernel Interface Explicit for Reliable CUDA Kernel Verification
Large language model (LLM) inference systems rely on CUDA kernels for core GPU computations, yet the interface between models and kernels is implicit and poorly specified. Models and kernels evolve independently and often make incompatible assumptions about tensor shapes and input sizes, leading to subtle memory bugs in CUDA kernels. These bugs can crash inference services, corrupt model weights, or be exploited by remote adversaries. Existing techniques either incur prohibitive runtime overhead, require specialized hardware, or fail to handle dynamic tensor shapes and variable kernel launch configurations, leaving the CUDA memory bugs largely unaddressed. This paper presents M2K, a fully automated framework that makes the model-kernel interface explicit and leverages it to detect memory bugs in CUDA kernels used in LLM inference systems. M2K consists of two components. HFProbe traces model execution without GPU hardware, classifies kernel arguments into model-fixed and user-variable, and emits symbolic constraints that capture the interface. cuK-LEE then performs symbolic execution on CUDA kernels to pinpoint memory bugs under the interface constraints, modeling tensors as disjoint memory regions and treating thread identifiers symbolically to scale to thousands of threads. In the evaluation, M2K discovers 181 previously unknown bugs in real LLM inference systems, while producing only nine false positives, demonstrating its effectiveness.
- Open Access
- 11 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 368-395
EPIC: Abstraction and Polymorphism of In-Network Collectives on Ethernet
- 11 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 368-395
Open AccessEPIC: Abstraction and Polymorphism of In-Network Collectives on Ethernet
- Open Access
- 11 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 334-352
Turbo: Efficiently Serving Long-Context Large Language Models with In-Network Aggregation
- 11 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 334-352
Open AccessTurbo: Efficiently Serving Long-Context Large Language Models with In-Network Aggregation
- Open Access
- 26 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 564-585
Multipath Collective Communication Beyond Scale-up Networks in GPU Clouds
- 26 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 564-585
Open AccessMultipath Collective Communication Beyond Scale-up Networks in GPU Clouds
- Open Access
- 26 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 862-878
LCMP: Distributed Long-Haul Cost-Aware Multi-Path Routing for Inter-Datacenter RDMA Networks
RDMA-empowered cloud services are gradually deployed across datacenters (DCs) with multiple paths, which exhibit new properties of path asymmetry, delayed congestion signals, and simultaneous flow routing collisions, and further fail existing routing methods. We present LCMP, a distributed long-haul cost-aware multi-path routing framework that aims to place RDMA flows on multiple inter-DC paths, achieving low-cost, low-latency, and congestion-responsive transmission. LCMP combines a control-plane path-quality score with compact on-switch congestion signals, where the former unifies quality assessment for asymmetric paths and the latter enables responsive reaction to path congestion. LCMP further resolves the simultaneous flow decision collision problem by filtering high-cost candidates, and performing a diversity-preserving hash inside the reduced set. On an 8-DC testbed, LCMP reduces median and tail FCT slowdown by up to 76% and 64%, respectively compared to state-of-the-art (SOTA) baselines. And large-scale NS-3 simulations under the 2000 km inter-DC scenario confirm similar improvements.
- 26 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 862-878
Open AccessLCMP: Distributed Long-Haul Cost-Aware Multi-Path Routing for Inter-Datacenter RDMA Networks
RDMA-empowered cloud services are gradually deployed across datacenters (DCs) with multiple paths, which exhibit new properties of path asymmetry, delayed congestion signals, and simultaneous flow routing collisions, and further fail existing routing methods. We present LCMP, a distributed long-haul cost-aware multi-path routing framework that aims to place RDMA flows on multiple inter-DC paths, achieving low-cost, low-latency, and congestion-responsive transmission. LCMP combines a control-plane path-quality score with compact on-switch congestion signals, where the former unifies quality assessment for asymmetric paths and the latter enables responsive reaction to path congestion. LCMP further resolves the simultaneous flow decision collision problem by filtering high-cost candidates, and performing a diversity-preserving hash inside the reduced set. On an 8-DC testbed, LCMP reduces median and tail FCT slowdown by up to 76% and 64%, respectively compared to state-of-the-art (SOTA) baselines. And large-scale NS-3 simulations under the 2000 km inter-DC scenario confirm similar improvements.
- 27 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Networking
- Vol. 34, 3947-3962
INARouting: Efficient Multi-Job Routing Optimization for Hierarchical In-Network Aggregation
In-network aggregation (INA) has emerged as a key technology to alleviate communication bottlenecks in large-scale distributed training, but its performance is often hindered by suboptimal routing. Existing INA-aware routing algorithms suffer from certain limitations: they either lack a global, multi-job coordination mechanism, or operate on incomplete network models that ignore key hardware constraints such as switch processing capacity. These deficiencies lead to network congestion and inefficient resource utilization, ultimately undermining the full potential of INA. To address these challenges, we present INARouting, a novel framework that holistically solves the multi-job hierarchical aggregation routing problem. We propose TINA, a hierarchical aggregation protocol that supports multi-job in-network aggregation. To address different deployment scenarios, we develop two variants: INARouting-Opt that provides optimal solutions for moderate-scale networks, and INARouting-Relax, a fast and effective heuristic using LP-relaxation and a greedy score-based rounding algorithm for large-scale deployments. Through extensive experiments on various scales of Fat-Tree and Spine-Leaf topologies, we demonstrate that INARouting significantly outperforms state-of-the-art methods. INARouting- Opt achieves provably optimal solutions, reducing average job completion time by up to 56% compared to existing methods. Meanwhile, INARouting-Relax outperforms existing algorithms while being 5× faster in solving time, enabling efficient routing in large-scale, dynamic environments.
- 27 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Networking
- Vol. 34, 3947-3962
INARouting: Efficient Multi-Job Routing Optimization for Hierarchical In-Network Aggregation
In-network aggregation (INA) has emerged as a key technology to alleviate communication bottlenecks in large-scale distributed training, but its performance is often hindered by suboptimal routing. Existing INA-aware routing algorithms suffer from certain limitations: they either lack a global, multi-job coordination mechanism, or operate on incomplete network models that ignore key hardware constraints such as switch processing capacity. These deficiencies lead to network congestion and inefficient resource utilization, ultimately undermining the full potential of INA. To address these challenges, we present INARouting, a novel framework that holistically solves the multi-job hierarchical aggregation routing problem. We propose TINA, a hierarchical aggregation protocol that supports multi-job in-network aggregation. To address different deployment scenarios, we develop two variants: INARouting-Opt that provides optimal solutions for moderate-scale networks, and INARouting-Relax, a fast and effective heuristic using LP-relaxation and a greedy score-based rounding algorithm for large-scale deployments. Through extensive experiments on various scales of Fat-Tree and Spine-Leaf topologies, we demonstrate that INARouting significantly outperforms state-of-the-art methods. INARouting- Opt achieves provably optimal solutions, reducing average job completion time by up to 56% compared to existing methods. Meanwhile, INARouting-Relax outperforms existing algorithms while being 5× faster in solving time, enabling efficient routing in large-scale, dynamic environments.
- 27 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Networking
- Vol. 34, 3766-3781
Virtual Slicing: Achieving Control Plane Availability and Traffic Engineering Efficiency in Data Centers
Many proposals have demonstrated the efficiency advantages of software-defined networking (SDN) in managing data center networks. Common practices employ centralized traffic engineering (TE) in the SDN control plane to optimize load balancing and throughput. Meanwhile, for high availability purposes, the control plane is partitioned to ensure the impact of a single faulty controller is contained. However, the interaction between these two aspects is often overlooked. In particular, we show that the current control plane partitioning approach leads to imbalanced link loads and degraded application performance. To address this issue, we propose virtual slicing, a new control plane partitioning scheme. Virtual slicing achieves desirable traffic engineering performance while retaining the availability guarantees from the current approach. Virtual slicing is implemented and evaluated with real-world and synthetic traffic traces on production spine-free data center networks. Results show that virtual slicing reduces tail link utilizations by up to 28.4%, and improves flow completion times by up to 36%.
- 27 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Networking
- Vol. 34, 3766-3781
Virtual Slicing: Achieving Control Plane Availability and Traffic Engineering Efficiency in Data Centers
Many proposals have demonstrated the efficiency advantages of software-defined networking (SDN) in managing data center networks. Common practices employ centralized traffic engineering (TE) in the SDN control plane to optimize load balancing and throughput. Meanwhile, for high availability purposes, the control plane is partitioned to ensure the impact of a single faulty controller is contained. However, the interaction between these two aspects is often overlooked. In particular, we show that the current control plane partitioning approach leads to imbalanced link loads and degraded application performance. To address this issue, we propose virtual slicing, a new control plane partitioning scheme. Virtual slicing achieves desirable traffic engineering performance while retaining the availability guarantees from the current approach. Virtual slicing is implemented and evaluated with real-world and synthetic traffic traces on production spine-free data center networks. Results show that virtual slicing reduces tail link utilizations by up to 28.4%, and improves flow completion times by up to 36%.
- 12 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Parallel and Distributed Systems
- Vol. 37 (4) , 920-933
HarmonyCache: Scalable In-Network Cache With Read-Write Separation
"In key-value storage systems, a small number of hot items account for most traffic. Skewed workloads lead to load imbalance among servers, and those holding hotspots become system bottlenecks, degrading overall performance. Recent studies show that load imbalance can be eliminated by deploying a small, fast cache node in front of back-end servers, i.e., in-network cache. Programmable switches enable placing such caches on switches where traffic must pass. Although existing in-network cache schemes effectively balance loads in large-scale storage systems, they perform poorly under write-intensive workloads and lose scalability with growing clients due to imbalanced cache nodes. This paper introduces HarmonyCache, a scalable, high-performance in-network cache system that supports write-back. HarmonyCache employs cache replication and read-write separation: only one cache node handles write requests, while others serve reads only. To achieve scalability and minimize coherence overhead, HarmonyCache proposes an adaptive cache replication scheme to determine where and how many replicas to deploy. In addition, we design heterogeneous in-network caches using different switch resources and propose a hybrid caching scheme. Prototype and extensive experiments show that HarmonyCache significantly improves throughput under various access patterns (read/write-intensive), achieving up to 7.6× throughput gain over state-of-the-art solutions under skewed write-intensive workloads.
- 12 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Parallel and Distributed Systems
- Vol. 37 (4) , 920-933
HarmonyCache: Scalable In-Network Cache With Read-Write Separation
"In key-value storage systems, a small number of hot items account for most traffic. Skewed workloads lead to load imbalance among servers, and those holding hotspots become system bottlenecks, degrading overall performance. Recent studies show that load imbalance can be eliminated by deploying a small, fast cache node in front of back-end servers, i.e., in-network cache. Programmable switches enable placing such caches on switches where traffic must pass. Although existing in-network cache schemes effectively balance loads in large-scale storage systems, they perform poorly under write-intensive workloads and lose scalability with growing clients due to imbalanced cache nodes. This paper introduces HarmonyCache, a scalable, high-performance in-network cache system that supports write-back. HarmonyCache employs cache replication and read-write separation: only one cache node handles write requests, while others serve reads only. To achieve scalability and minimize coherence overhead, HarmonyCache proposes an adaptive cache replication scheme to determine where and how many replicas to deploy. In addition, we design heterogeneous in-network caches using different switch resources and propose a hybrid caching scheme. Prototype and extensive experiments show that HarmonyCache significantly improves throughput under various access patterns (read/write-intensive), achieving up to 7.6× throughput gain over state-of-the-art solutions under skewed write-intensive workloads.
- Open Access
- 30 November 2025
- Published by Association for Computing Machinery (ACM)
- p. 3-9
PRISM: A Workload-Predictive Job Scheduler for Clusters with In-Network Aggregation Capability
- 30 November 2025
- Published by Association for Computing Machinery (ACM)
- p. 3-9
Open AccessPRISM: A Workload-Predictive Job Scheduler for Clusters with In-Network Aggregation Capability