Siwei Ji
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Siwei Ji has a total of 11 co-authors. Here are the top co-authors by number of shared publications:
- Bo Lei
China Telecom, China
1 shared publication - Guozhen Dong
China Telecom Corporation Limited Research Institute, Beijing, China
1 shared publication - Jing Tang
China Telecom Research Institute, Beijing, China
1 shared publication - Nan Wu
China Telecom Corporation Limited Research Institute, Beijing, China
1 shared publication - Wenjuan Xing
China Telecom Corporation Limited Research Institute, Beijing, China
1 shared publication - Xiaoting Ma
China Telecom Research Institute, Beijing, China
1 shared publication
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.
- Open Access
- 18 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 1431-1435
A Computing Power Method Based on EVPN-VXLAN Extension in Data Center Networks
Data center networks need to deliver various cloud services to large number of users, necessitating efficient and flexible management of network resources. However, the widely utilized EVPN-VXLAN (Ethernet Virtual Private Network - Virtual Extensible Local Area Network) technology for network virtualization currently lacks support for computing power resource scheduling within the overlay network. While computing power services can conserve network bandwidth and reduce latency for high-demand applications, they encounter challenges related to end-to-end delivery, path selection, and resource allocation in the EVPN-VXLAN data center network. This paper proposes a novel computing power scheduling method based on EVPN-VXLAN that defines a new EVPN routing type to facilitate the delivery of multi-resource information, which is then utilized for computing power routing selection and forwarding. The proposed method enhances the performance and efficiency of computing power services within the data center EVPN-VXLAN network at the overlay layer while maintaining compatibility with existing EVPN-VXLAN technology.
- 18 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 1431-1435
Open AccessA Computing Power Method Based on EVPN-VXLAN Extension in Data Center Networks
Data center networks need to deliver various cloud services to large number of users, necessitating efficient and flexible management of network resources. However, the widely utilized EVPN-VXLAN (Ethernet Virtual Private Network - Virtual Extensible Local Area Network) technology for network virtualization currently lacks support for computing power resource scheduling within the overlay network. While computing power services can conserve network bandwidth and reduce latency for high-demand applications, they encounter challenges related to end-to-end delivery, path selection, and resource allocation in the EVPN-VXLAN data center network. This paper proposes a novel computing power scheduling method based on EVPN-VXLAN that defines a new EVPN routing type to facilitate the delivery of multi-resource information, which is then utilized for computing power routing selection and forwarding. The proposed method enhances the performance and efficiency of computing power services within the data center EVPN-VXLAN network at the overlay layer while maintaining compatibility with existing EVPN-VXLAN technology.