Ziheng Xu
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
Ziheng Xu has a total of 26 co-authors. Here are the top co-authors by number of shared publications:
- Wei Quan
School of Electronic and Information Engineering, Beijing Jiaotong University, Beijing, China
6 shared publications - Hongke Zhang
Beijing Jiaotong University
5 shared publications - Nan Cheng
Xidian University
5 shared publications - Deyun Gao
National Engineering Research Center of Advanced Network Technologies, School of Electronic and Information Engineering, Beijing Jiaotong University, Beijing, China
4 shared publications - Mingyuan Liu
School of Electronic and Information Engineering, Beijing Jiaotong University, Beijing, China
4 shared publications - Kang Liu
School of Electronic and Information Engineering, Beijing Jiaotong University, Beijing, China
2 shared publications
Latest Publications
- Open Access
- 11 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 1828-1832
LST-Sim: An Efficient Simulation Platform for Large-Scale Model Training
Existing model training simulators fail to support heterogeneous GPU architectures and various intra-server high-speed interconnect protocols (e.g., NVLink, HCCS), while suffering from low efficiency in full-packet-level network simulation. In this paper, we present LST-Sim, an efficient simulator designed for large-scale model training, balancing simulation accuracy and runtime efficiency. First, the simulator takes into account the impacts of computational heterogeneity across different GPU models and inter-GPU interconnection bandwidth on training latency, and constructs a theoretical model. Furthermore, we propose a domain-based traffic management mechanism that integrates theoretical analysis with network simulation to handle different types of communication traffic, thereby improving simulation efficiency. The effectiveness of LST-Sim is well validated on the Ascend 910B training cluster. Its simulation results are in close agreement with real-world measurements. Meanwhile, the domain-based traffic management mechanism significantly reduces simulation runtime by approximately 75% while maintaining high fidelity.
- 11 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 1828-1832
Open AccessLST-Sim: An Efficient Simulation Platform for Large-Scale Model Training
Existing model training simulators fail to support heterogeneous GPU architectures and various intra-server high-speed interconnect protocols (e.g., NVLink, HCCS), while suffering from low efficiency in full-packet-level network simulation. In this paper, we present LST-Sim, an efficient simulator designed for large-scale model training, balancing simulation accuracy and runtime efficiency. First, the simulator takes into account the impacts of computational heterogeneity across different GPU models and inter-GPU interconnection bandwidth on training latency, and constructs a theoretical model. Furthermore, we propose a domain-based traffic management mechanism that integrates theoretical analysis with network simulation to handle different types of communication traffic, thereby improving simulation efficiency. The effectiveness of LST-Sim is well validated on the Ascend 910B training cluster. Its simulation results are in close agreement with real-world measurements. Meanwhile, the domain-based traffic management mechanism significantly reduces simulation runtime by approximately 75% while maintaining high fidelity.
- 11 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Mobile Computing
- Vol. 24 (8) , 7312-7327
HarmonyPath: Fine-Grained Flexible Multipath Transmission for Mobile Differentiated Services
The surge in mobile application services has led to diversified traffic and increased demands on network resources. Traditional multipath algorithms, designed for resource integration through subflow scheduling across paths, struggle with disharmonious transmission caused by terminal mobility and differentiated path resources. Especially when differentiated services are transmitted concurrently, disharmonious transmission can give rise to resource contention, causing a large number of subflows to congest a single path and leading to performance degradation. To mitigate these challenges, this paper introduces HarmonyPath, a fine-grained flexible multipath transmission mechanism that can ensure harmonious resource occupation. Specifically, HarmonyPath firstly employs an in-band telemetry protocol to gather path resource information, generating a network resource distribution map. Based on this map, it flexibly allocates path resources according to the network resource distribution and service requirements. Then, HarmonyPath establishes a collaborative matching model for service demands and path resources. Through matrix transformation and calculation, it rapidly generates and deploys the scheduling strategy. To further alleviate service contention, HarmonyPath employs heuristic algorithms to optimize the scheduling strategy and achieve precise multipath transmission. Experiments demonstrate that HarmonyPath surpasses traditional algorithms in the multipath transmission of differentiated services, offering flexible service resource guarantees and enhancing network resource utilization efficiency.
- 11 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Mobile Computing
- Vol. 24 (8) , 7312-7327
HarmonyPath: Fine-Grained Flexible Multipath Transmission for Mobile Differentiated Services
The surge in mobile application services has led to diversified traffic and increased demands on network resources. Traditional multipath algorithms, designed for resource integration through subflow scheduling across paths, struggle with disharmonious transmission caused by terminal mobility and differentiated path resources. Especially when differentiated services are transmitted concurrently, disharmonious transmission can give rise to resource contention, causing a large number of subflows to congest a single path and leading to performance degradation. To mitigate these challenges, this paper introduces HarmonyPath, a fine-grained flexible multipath transmission mechanism that can ensure harmonious resource occupation. Specifically, HarmonyPath firstly employs an in-band telemetry protocol to gather path resource information, generating a network resource distribution map. Based on this map, it flexibly allocates path resources according to the network resource distribution and service requirements. Then, HarmonyPath establishes a collaborative matching model for service demands and path resources. Through matrix transformation and calculation, it rapidly generates and deploys the scheduling strategy. To further alleviate service contention, HarmonyPath employs heuristic algorithms to optimize the scheduling strategy and achieve precise multipath transmission. Experiments demonstrate that HarmonyPath surpasses traditional algorithms in the multipath transmission of differentiated services, offering flexible service resource guarantees and enhancing network resource utilization efficiency.
- 25 December 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Mobile Computing
- Vol. 23 (8) , 8274-8288
DOFMS: DRL-Based Out-of-Order Friendly Multipath Scheduling in Mobile Heterogeneous Networks
Multipath transmission brings strong bandwidth aggregation capability for services in wireless networks. Nonetheless, the heterogeneous nature of paths and the motion of terminals results in varying transmission delays, leading to out-of-order (OFO) delivery and transmission quality decrease. Traditional algorithms, due to their solid operating logic, fail to strike a balance between high bandwidth and low OFO extent. Recent studies have focused on utilizing learning algorithms to find a multi-metric joint optimal transmission strategy. In light of this, this paper proposes a framework called DRL-based OFO-Friendly Multipath Scheduling (DOFMS) to ensure high bandwidth and low OFO extent transmission in mobile heterogeneous networks. In particular, the framework introduces a novel OFO evaluation index to assess the degree of OFO more accurately. To achieve elastic scheduling, the framework employs the Double Deep Q Network (DDQN) to dynamically regulate the scheduling ratio. Recognizing the dynamic and unpredictable nature of path delays, an asynchronous module is introduced to enhance learning accuracy. Experimental results demonstrate that the DOFMS reduces the OFO rate by 25% compared to traditional bandwidth aggregation algorithms, while maintaining low bandwidth and packet loss rates. Furthermore, compared to conventional OFO avoidance algorithms, the DOFMS improves bandwidth by 4% and reduces fluctuation by 90%.
- 25 December 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Mobile Computing
- Vol. 23 (8) , 8274-8288
DOFMS: DRL-Based Out-of-Order Friendly Multipath Scheduling in Mobile Heterogeneous Networks
Multipath transmission brings strong bandwidth aggregation capability for services in wireless networks. Nonetheless, the heterogeneous nature of paths and the motion of terminals results in varying transmission delays, leading to out-of-order (OFO) delivery and transmission quality decrease. Traditional algorithms, due to their solid operating logic, fail to strike a balance between high bandwidth and low OFO extent. Recent studies have focused on utilizing learning algorithms to find a multi-metric joint optimal transmission strategy. In light of this, this paper proposes a framework called DRL-based OFO-Friendly Multipath Scheduling (DOFMS) to ensure high bandwidth and low OFO extent transmission in mobile heterogeneous networks. In particular, the framework introduces a novel OFO evaluation index to assess the degree of OFO more accurately. To achieve elastic scheduling, the framework employs the Double Deep Q Network (DDQN) to dynamically regulate the scheduling ratio. Recognizing the dynamic and unpredictable nature of path delays, an asynchronous module is introduced to enhance learning accuracy. Experimental results demonstrate that the DOFMS reduces the OFO rate by 25% compared to traditional bandwidth aggregation algorithms, while maintaining low bandwidth and packet loss rates. Furthermore, compared to conventional OFO avoidance algorithms, the DOFMS improves bandwidth by 4% and reduces fluctuation by 90%.
- 25 August 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 168-171
Multi-level Feature Aggregation Network for Person Re-identification
Current person re-identification works focus on the deep features, ignoring the loss of detailed features due to downsampling operations. We propose a parallel structured Multi-level Feature Aggregation network (MFA-Net), which not only aggregates the information of local features within the same level as that of global features but also aggregates the features between different levels in an orderly manner as a way to mine the overlooked features. Furthermore, experiments on three datasets Market1501, DukeMTMC-RelD and MSMT17, show that our MFA-Net can further mine the detailed features, enhance the feature representation, and achieve state-of-the-art results on three benchmark datasets.
- 25 August 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 168-171
Multi-level Feature Aggregation Network for Person Re-identification
Current person re-identification works focus on the deep features, ignoring the loss of detailed features due to downsampling operations. We propose a parallel structured Multi-level Feature Aggregation network (MFA-Net), which not only aggregates the information of local features within the same level as that of global features but also aggregates the features between different levels in an orderly manner as a way to mine the overlooked features. Furthermore, experiments on three datasets Market1501, DukeMTMC-RelD and MSMT17, show that our MFA-Net can further mine the detailed features, enhance the feature representation, and achieve state-of-the-art results on three benchmark datasets.
- 23 May 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Vehicular Technology
- Vol. 72 (10) , 13575-13590
Reliable PPO-Based Concurrent Multipath Transfer for Time-Sensitive Applications
Time-sensitive applications, e.g. , Internet of vehicles applications and tactile Internet applications, put forward low latency and high throughput requirements for communication. However, due to the dynamical network conditions, frequent retransmission will produce undesirable latency and reduce throughput. This article proposes a reliable proximal policy optimization (PPO)-based concurrent multipath transfer mechanism (RpCMT) that can reduce retransmission probability, as a consequence, achieves low latency and high throughput. In particular, RpCMT employs in-network telemetry (INT)-based network-aware transmission control algorithm to make dynamical multipath transfer control depending on the link conditions. To generate precise transfer control policy, we formulate multipath flow scheduling as a minimization problem that considers both the link bandwidth utilization and transmission queuing length. Then, we propose a PPO-based multipath transfer policy updating algorithm to obtain the optimized control policy. Finally, we build a P4-based prototype system in which RpCMT is implemented to make experimental analysis. Experimental results show that RpCMT outperforms the state-of-the-art solutions, e.g. , weighted random routing (W-Random) and link weighted based ECMP (W-ECMP), in terms of the retransmission rate, throughput, and round-trip time.
- 23 May 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Vehicular Technology
- Vol. 72 (10) , 13575-13590
Reliable PPO-Based Concurrent Multipath Transfer for Time-Sensitive Applications
Time-sensitive applications, e.g. , Internet of vehicles applications and tactile Internet applications, put forward low latency and high throughput requirements for communication. However, due to the dynamical network conditions, frequent retransmission will produce undesirable latency and reduce throughput. This article proposes a reliable proximal policy optimization (PPO)-based concurrent multipath transfer mechanism (RpCMT) that can reduce retransmission probability, as a consequence, achieves low latency and high throughput. In particular, RpCMT employs in-network telemetry (INT)-based network-aware transmission control algorithm to make dynamical multipath transfer control depending on the link conditions. To generate precise transfer control policy, we formulate multipath flow scheduling as a minimization problem that considers both the link bandwidth utilization and transmission queuing length. Then, we propose a PPO-based multipath transfer policy updating algorithm to obtain the optimized control policy. Finally, we build a P4-based prototype system in which RpCMT is implemented to make experimental analysis. Experimental results show that RpCMT outperforms the state-of-the-art solutions, e.g. , weighted random routing (W-Random) and link weighted based ECMP (W-ECMP), in terms of the retransmission rate, throughput, and round-trip time.
- 27 October 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Communications Surveys & Tutorials
- Vol. 25 (1) , 762-790
AI-Driven Packet Forwarding With Programmable Data Plane: A Survey
The existing packet forwarding technology cannot meet the increasing requirements of Internet development due to its rigid framework. Application of artificial intelligence (AI) for efficient packet forwarding is gaining more and more interest as a new direction. Recently, the explosive development of programmable data plane (PDP) has provided a potential impetus to packet forwarding driven by AI. Therefore, this paper presents a survey on the recent research in AI-driven packet forwarding with PDP. First, we describe two of the most representative frameworks of the packet forwarding, i.e., the traditional AI-driven forwarding framework and the new one assisted by the PDP. Then, we focus on capacity of the packet forwarding under the two frameworks in four measures: delay, throughput, security, and reliability. For each measure, we organize the content with the evolution from simple packet forwarding, to packet forwarding capacity enhancement with the assistance of AI, to the latest research on AI-driven packet forwarding supported by the PDP. Finally, we identify three directions in the development of AI-driven packet forwarding, and highlight the challenges and issues in future research.
- 27 October 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Communications Surveys & Tutorials
- Vol. 25 (1) , 762-790
AI-Driven Packet Forwarding With Programmable Data Plane: A Survey
The existing packet forwarding technology cannot meet the increasing requirements of Internet development due to its rigid framework. Application of artificial intelligence (AI) for efficient packet forwarding is gaining more and more interest as a new direction. Recently, the explosive development of programmable data plane (PDP) has provided a potential impetus to packet forwarding driven by AI. Therefore, this paper presents a survey on the recent research in AI-driven packet forwarding with PDP. First, we describe two of the most representative frameworks of the packet forwarding, i.e., the traditional AI-driven forwarding framework and the new one assisted by the PDP. Then, we focus on capacity of the packet forwarding under the two frameworks in four measures: delay, throughput, security, and reliability. For each measure, we organize the content with the evolution from simple packet forwarding, to packet forwarding capacity enhancement with the assistance of AI, to the latest research on AI-driven packet forwarding supported by the PDP. Finally, we identify three directions in the development of AI-driven packet forwarding, and highlight the challenges and issues in future research.
- 6 August 2022
- Published by Springer Nature in Peer-to-Peer Networking and Applications
- Vol. 15 (6) , 2744-2756
Deep reinforcement learning-based fountain coding for concurrent multipath transfer in high-speed railway networks
- 6 August 2022
- Published by Springer Nature in Peer-to-Peer Networking and Applications
- Vol. 15 (6) , 2744-2756
Deep reinforcement learning-based fountain coding for concurrent multipath transfer in high-speed railway networks
- 16 May 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 4655-4660
PPO-based Reliable Concurrent Transmission Control for Telemedicine Real-time Services
Telemedicine services put forward high transmission demands for network transmission, such as low latency and high throughput. However, telemedicine services suffer undesirable latency due to the high re-transmission probability caused by congestion and queuing. To reduce the probability of re-transmission, this paper firstly proposes in-band network telemetry (INT)-based delay-guaranteed transmission framework (IDTF) to make concurrent transmission control. In IDTF, we propose a proximal policy optimization (PPO)-based adaptive multipath concurrent flow scheduling algorithm (PAMA) for control policy adjustment. In detail, PAMA makes a joint minimum optimization with flow scheduling and network resource management to reduce the probability of congestion and long-time queuing. Finally, we implement extensive simulations on a programming protocol-independent packet processors (P4)-based programmable network platform to perform performance analysis. Simulation results show that PAMA outperforms existing classical algorithms in re-transmission rate, round-trip time, and throughput.
- 16 May 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 4655-4660
PPO-based Reliable Concurrent Transmission Control for Telemedicine Real-time Services
Telemedicine services put forward high transmission demands for network transmission, such as low latency and high throughput. However, telemedicine services suffer undesirable latency due to the high re-transmission probability caused by congestion and queuing. To reduce the probability of re-transmission, this paper firstly proposes in-band network telemetry (INT)-based delay-guaranteed transmission framework (IDTF) to make concurrent transmission control. In IDTF, we propose a proximal policy optimization (PPO)-based adaptive multipath concurrent flow scheduling algorithm (PAMA) for control policy adjustment. In detail, PAMA makes a joint minimum optimization with flow scheduling and network resource management to reduce the probability of congestion and long-time queuing. Finally, we implement extensive simulations on a programming protocol-independent packet processors (P4)-based programmable network platform to perform performance analysis. Simulation results show that PAMA outperforms existing classical algorithms in re-transmission rate, round-trip time, and throughput.