Longxiang Wang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Interests
Longxiang Wang has a total of 26 co-authors. Here are the top co-authors by number of shared publications:
- Xingjun Zhang
Xi’an Jiaotong University
4 shared publications - Jia Wei
Tsinghua University
3 shared publications - Xiaoshe Dong
School of Computer Science and Technology, Xi'an Jiaotong university
3 shared publications - Yufei Wang
School of Electronic and Information Engineering, Xi’an Jiaotong University, Xi’an, China
3 shared publications - Witold Pedrycz
University of Alberta
2 shared publications - Bin Liang
Xian Shiyou University, Xian, China
1 shared publication
Latest Publications
- 1 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Emerging Topics in Computational Intelligence
- Vol. 10 (3) , 2291-2300
Homophone-Based Chinese Natural Language Data Augmentation
Data, devices, and algorithms drive progress in artificial intelligence, and high quality data is essential for model training. However, collecting and labeling such data is costly. Inspired by computer vision, natural language data augmentation has attracted attention as a way to expand corpora through random insertion, deletion, synonym replacement, swapping, and large language model based generation. Most Chinese augmentation methods still follow English style strategies and fail to exploit Chinese specific linguistic properties, which limits their usability and effectiveness. Chinese users often make homophone errors in instant messaging programs such as QQ, WeChat, and WhatsApp. However, readers can still understand the intended meaning from context, background, and language habits. We call this phenomenon homophone consistency. We exploit Chinese homophone consistency for Chinese natural language data augmentation and design a Homophone-based Chinese Augmentation. HCA converts a sentence into a pinyin queue and then converts the pinyin queue back into a sentence using a dictionary based dynamic programming algorithm for pinyin identification. Words in the original sentence are replaced with homophones with a specified probability. HCA broadens Chinese augmentation using Chinese features, enriches the Chinese data augmentation strategy space, and provides a reference for designing language specific augmentation methods for other languages. We fine tune Chinese BERT on four datasets of different sizes and evaluate test accuracy. HCA improves test accuracy by up to 37.62% when the dataset is small and improves it by 3.42% on average. We also conduct ablation studies to analyze the effects of key parameters on training outcomes.
- 1 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Emerging Topics in Computational Intelligence
- Vol. 10 (3) , 2291-2300
Homophone-Based Chinese Natural Language Data Augmentation
Data, devices, and algorithms drive progress in artificial intelligence, and high quality data is essential for model training. However, collecting and labeling such data is costly. Inspired by computer vision, natural language data augmentation has attracted attention as a way to expand corpora through random insertion, deletion, synonym replacement, swapping, and large language model based generation. Most Chinese augmentation methods still follow English style strategies and fail to exploit Chinese specific linguistic properties, which limits their usability and effectiveness. Chinese users often make homophone errors in instant messaging programs such as QQ, WeChat, and WhatsApp. However, readers can still understand the intended meaning from context, background, and language habits. We call this phenomenon homophone consistency. We exploit Chinese homophone consistency for Chinese natural language data augmentation and design a Homophone-based Chinese Augmentation. HCA converts a sentence into a pinyin queue and then converts the pinyin queue back into a sentence using a dictionary based dynamic programming algorithm for pinyin identification. Words in the original sentence are replaced with homophones with a specified probability. HCA broadens Chinese augmentation using Chinese features, enriches the Chinese data augmentation strategy space, and provides a reference for designing language specific augmentation methods for other languages. We fine tune Chinese BERT on four datasets of different sizes and evaluate test accuracy. HCA improves test accuracy by up to 37.62% when the dataset is small and improves it by 3.42% on average. We also conduct ablation studies to analyze the effects of key parameters on training outcomes.
- 29 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computers
- Vol. 75 (3) , 1209-1223
Dual-Pronged Deep Learning Preprocessing on Heterogeneous Platforms With CPU, Accelerator and CSD
For image-related deep learning tasks, the first step often involves reading data from external storage and performing preprocessing on the CPU. As accelerator speed increases and the number of single compute node accelerators increases, the computing and data transfer capabilities gap between accelerators and CPUs gradually increases. Data reading and preprocessing become progressively the bottleneck of these tasks. Our work, DDLP, addresses the data computing and transfer bottleneck of deep learning preprocessing using Computable Storage Devices (CSDs). DDLP allows the CPU and CSD to efficiently parallelize preprocessing from both ends of the datasets, respectively. To this end, we propose two adaptive dynamic selection strategies to make DDLP control the accelerator to automatically read data from different sources. The two strategies trade-off between consistency and efficiency. DDLP achieves sufficient computational overlap between CSD data preprocessing and CPU preprocessing, accelerator computation, and accelerator data reading. In addition, DDLP leverages direct storage technology to enable efficient SSD-to-accelerator data transfer. In addition, DDLP reduces the use of expensive CPU and DRAM resources with more energy-efficient CSDs, alleviating preprocessing bottlenecks while significantly reducing power consumption. Extensive experimental results show that DDLP can improve learning speed by up to 23.5% on ImageNet Dataset while reducing energy consumption by 19.7% and CPU and DRAM usage by 37.6%. DDLP also improves the learning speed by up to 27.6% on the Cifar-10 dataset.
- 29 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computers
- Vol. 75 (3) , 1209-1223
Dual-Pronged Deep Learning Preprocessing on Heterogeneous Platforms With CPU, Accelerator and CSD
For image-related deep learning tasks, the first step often involves reading data from external storage and performing preprocessing on the CPU. As accelerator speed increases and the number of single compute node accelerators increases, the computing and data transfer capabilities gap between accelerators and CPUs gradually increases. Data reading and preprocessing become progressively the bottleneck of these tasks. Our work, DDLP, addresses the data computing and transfer bottleneck of deep learning preprocessing using Computable Storage Devices (CSDs). DDLP allows the CPU and CSD to efficiently parallelize preprocessing from both ends of the datasets, respectively. To this end, we propose two adaptive dynamic selection strategies to make DDLP control the accelerator to automatically read data from different sources. The two strategies trade-off between consistency and efficiency. DDLP achieves sufficient computational overlap between CSD data preprocessing and CPU preprocessing, accelerator computation, and accelerator data reading. In addition, DDLP leverages direct storage technology to enable efficient SSD-to-accelerator data transfer. In addition, DDLP reduces the use of expensive CPU and DRAM resources with more energy-efficient CSDs, alleviating preprocessing bottlenecks while significantly reducing power consumption. Extensive experimental results show that DDLP can improve learning speed by up to 23.5% on ImageNet Dataset while reducing energy consumption by 19.7% and CPU and DRAM usage by 37.6%. DDLP also improves the learning speed by up to 27.6% on the Cifar-10 dataset.
- 31 May 2024
- Published by Elsevier in Future Generation Computer Systems
- Vol. 160, 140-153
Flimm: Foreground traffic aware data migration manager for distributed storage system
- 31 May 2024
- Published by Elsevier in Future Generation Computer Systems
- Vol. 160, 140-153
Flimm: Foreground traffic aware data migration manager for distributed storage system
- Open Access
- 14 December 2023
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 20 (4) , 1-25
Fastensor: Optimise the Tensor I/O Path from SSD to GPU for Deep Learning Training
In recent years, benefiting from the increase in model size and complexity, deep learning has achieved tremendous success in computer vision (CV) and (NLP). Training deep learning models using accelerators such as GPUs often requires much iterative data to be transferred from NVMe SSD to GPU memory. Much recent work has focused on data transfer during the pre-processing phase and has introduced techniques such as multiprocessing and GPU Direct Storage (GDS) to accelerate it. However, tensor data during training (such as Checkpoints, logs, and intermediate feature maps), which is also time-consuming, is often transferred using traditional serial, long-I/O-path transfer methods. In this article, based on GDS technology, we built Fastensor, an efficient tool for tensor data transfer between the NVMe SSDs and GPUs. To achieve higher tensor data I/O throughput, we optimized the traditional data I/O process. We also proposed a data and runtime context-aware tensor I/O algorithm. Fastensor can select the most suitable data transfer tool for the current tensor from a candidate set of tools during model training. The optimal tool is derived from a dictionary generated by our adaptive exploration algorithm in the first few training iterations. We used Fastensor’s unified interface to test the read/write bandwidth and energy consumption of different transfer tools for different sizes of tensor blocks. We found that the execution efficiency of different tensor transfer tools is related to both the tensor block size and the runtime context. We then deployed Fastensor in the widely applicable Pytorch deep learning framework. We showed that Fastensor could perform superior in typical scenarios of model parameter saving and intermediate feature map transfer with the same hardware configuration. Fastensor achieves a 5.37x read performance improvement compared to torch.save () when used for model parameter saving. When used for intermediate feature map transfer, Fastensor can increase the supported training batch size by 20x, while the total read and write speed is increased by 2.96x compared to the torch I/O API.
- 14 December 2023
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 20 (4) , 1-25
Open AccessFastensor: Optimise the Tensor I/O Path from SSD to GPU for Deep Learning Training
In recent years, benefiting from the increase in model size and complexity, deep learning has achieved tremendous success in computer vision (CV) and (NLP). Training deep learning models using accelerators such as GPUs often requires much iterative data to be transferred from NVMe SSD to GPU memory. Much recent work has focused on data transfer during the pre-processing phase and has introduced techniques such as multiprocessing and GPU Direct Storage (GDS) to accelerate it. However, tensor data during training (such as Checkpoints, logs, and intermediate feature maps), which is also time-consuming, is often transferred using traditional serial, long-I/O-path transfer methods. In this article, based on GDS technology, we built Fastensor, an efficient tool for tensor data transfer between the NVMe SSDs and GPUs. To achieve higher tensor data I/O throughput, we optimized the traditional data I/O process. We also proposed a data and runtime context-aware tensor I/O algorithm. Fastensor can select the most suitable data transfer tool for the current tensor from a candidate set of tools during model training. The optimal tool is derived from a dictionary generated by our adaptive exploration algorithm in the first few training iterations. We used Fastensor’s unified interface to test the read/write bandwidth and energy consumption of different transfer tools for different sizes of tensor blocks. We found that the execution efficiency of different tensor transfer tools is related to both the tensor block size and the runtime context. We then deployed Fastensor in the widely applicable Pytorch deep learning framework. We showed that Fastensor could perform superior in typical scenarios of model parameter saving and intermediate feature map transfer with the same hardware configuration. Fastensor achieves a 5.37x read performance improvement compared to torch.save () when used for model parameter saving. When used for intermediate feature map transfer, Fastensor can increase the supported training batch size by 20x, while the total read and write speed is increased by 2.96x compared to the torch I/O API.
- 1 June 2022
- Published by Springer Nature in CCF Transactions on High Performance Computing
- Vol. 4 (2) , 211-230
Status, challenges and trends of data-intensive supercomputing
- 1 June 2022
- Published by Springer Nature in CCF Transactions on High Performance Computing
- Vol. 4 (2) , 211-230
Status, challenges and trends of data-intensive supercomputing
- Open Access
- 9 October 2021
- Published by MDPI AG in Future Internet
- Vol. 13 (10) , 261
An Intelligent TCP Congestion Control Method Based on Deep Q Network
To optimize the data migration performance between different supercomputing centers in China, we present TCP-DQN, which is an intelligent TCP congestion control method based on DQN (Deep Q network). The TCP congestion control process is abstracted as a partially observed Markov decision process. In this process, an agent is constructed to interact with the network environment. The agent adjusts the size of the congestion window by observing the characteristics of the network state. The network environment feeds back the reward to the agent, and the agent tries to maximize the expected reward in an episode. We designed a weighted reward function to balance the throughput and delay. Compared with traditional Q-learning, DQN uses double-layer neural networks and experience replay to reduce the oscillation problem that may occur in gradient descent. We implemented the TCP-DQN method and compared it with mainstream congestion control algorithms such as cubic, Highspeed and NewReno. The results show that the throughput of TCP-DQN can reach more than 2 times of the comparison method while the latency is close to the three compared methods.
- 9 October 2021
- Published by MDPI AG in Future Internet
- Vol. 13 (10) , 261
Open AccessAn Intelligent TCP Congestion Control Method Based on Deep Q Network
To optimize the data migration performance between different supercomputing centers in China, we present TCP-DQN, which is an intelligent TCP congestion control method based on DQN (Deep Q network). The TCP congestion control process is abstracted as a partially observed Markov decision process. In this process, an agent is constructed to interact with the network environment. The agent adjusts the size of the congestion window by observing the characteristics of the network state. The network environment feeds back the reward to the agent, and the agent tries to maximize the expected reward in an episode. We designed a weighted reward function to balance the throughput and delay. Compared with traditional Q-learning, DQN uses double-layer neural networks and experience replay to reduce the oscillation problem that may occur in gradient descent. We implemented the TCP-DQN method and compared it with mainstream congestion control algorithms such as cubic, Highspeed and NewReno. The results show that the throughput of TCP-DQN can reach more than 2 times of the comparison method while the latency is close to the three compared methods.
- 7 May 2021
- Published by Springer Nature in The Journal of Supercomputing
- Vol. 77 (12) , 14011-14038
A multi-dimensional double descending maximum padding priority algorithm for cloud data centers
- 7 May 2021
- Published by Springer Nature in The Journal of Supercomputing
- Vol. 77 (12) , 14011-14038
A multi-dimensional double descending maximum padding priority algorithm for cloud data centers
- Open Access
- 16 November 2019
- Published by MDPI AG in Electronics
- Vol. 8 (11) , 1357
Measurement and Analysis of SSD Reliability Data Based on Accelerated Endurance Test
In recent years, NAND Flash-based solid-state drives (SSDs) have become more widely used in data centers and consumer markets. Data centers generally choose to provide high-quality storage services by deploying a large number of SSDs, but there are no effective preventive measures to reduce the impact of SSD failures currently. Some existing studies have analyzed the relevant factors related to SSD failures from different angles, but the characteristics of reliability changes exhibited by SSD throughout the life cycle have not been explored in depth. On the other hand, although the 3D manufacturing process has increased the storage density of the SSD, the mutual influence between the flash units has also increased, resulting in severe degradation of the performance and lifetime of the SSD. Therefore, in order to fully understand the reliability varying process of SSD throughout the life cycle, we first designed an SSD lifetime endurance test method, then conducted the endurance test and collected the reliability data for the entire life cycle of the 3D TLC SSD in the laboratory environment with reference to the JEDEC standard. Through the analysis of experimental data and its statistical correlation, it is found that SSD will produce a large number of uncorrectable errors before reaching the endurance limit, and there will be a phenomenon of continuous high operating temperature, as well as showing some intrinsic relationships about SSD reliability data. The findings in this paper are valuable for identifying whether an SSD is going to fail.
- 16 November 2019
- Published by MDPI AG in Electronics
- Vol. 8 (11) , 1357
Open AccessMeasurement and Analysis of SSD Reliability Data Based on Accelerated Endurance Test
In recent years, NAND Flash-based solid-state drives (SSDs) have become more widely used in data centers and consumer markets. Data centers generally choose to provide high-quality storage services by deploying a large number of SSDs, but there are no effective preventive measures to reduce the impact of SSD failures currently. Some existing studies have analyzed the relevant factors related to SSD failures from different angles, but the characteristics of reliability changes exhibited by SSD throughout the life cycle have not been explored in depth. On the other hand, although the 3D manufacturing process has increased the storage density of the SSD, the mutual influence between the flash units has also increased, resulting in severe degradation of the performance and lifetime of the SSD. Therefore, in order to fully understand the reliability varying process of SSD throughout the life cycle, we first designed an SSD lifetime endurance test method, then conducted the endurance test and collected the reliability data for the entire life cycle of the 3D TLC SSD in the laboratory environment with reference to the JEDEC standard. Through the analysis of experimental data and its statistical correlation, it is found that SSD will produce a large number of uncorrectable errors before reaching the endurance limit, and there will be a phenomenon of continuous high operating temperature, as well as showing some intrinsic relationships about SSD reliability data. The findings in this paper are valuable for identifying whether an SSD is going to fail.