Xiaoshe Dong
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Xiaoshe Dong has a total of 54 co-authors. Here are the top co-authors by number of shared publications:
- Ziheng Wang
School of Computer Science and Technology, Xi'an Jiaotong university
11 shared publications - Qiang Wang
Network Information Center, Xi'an Jiaotong University
8 shared publications - Heng Chen
School of Computer Science and Technology, Xian Jiaotong University, Xian, China
7 shared publications - Yan Kang
School of Computer Science and Technology, Xian Jiaotong University, Xian, China
7 shared publications - Xingjun Zhang
Xi’an Jiaotong University
6 shared publications - Weiduo Chen
School of Computer Science and Technology, Xi’an Jiaotong University, Xi’an, China
5 shared publications
Latest Publications
- Open Access
- 2 November 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
TSE-IO: Two-Stage Extrapolative I/O Modeling and Tuning via Pairwise Learning
Collecting I/O data for parallel programs at large node scales is prohibitively expensive and has long been a major bottleneck for I/O performance modeling. Existing approaches typically depend on extensive training samples collected on large-scale systems, which incur high cost and generalize poorly across applications and environments. To address these challenges, we propose TSE-IO, a T wo- S tage framework for E xtrapolative I/O performance modeling and tuning. TSE-IO decomposes cross-scale performance prediction into two linked tasks: first, extrapolating application-level I/O behavior from small-scale measurements to the target scale. And second, predicting target-scale performance by jointly modeling the extrapolated behavior and tunable I/O-stack configurations. For the second stage, TSE-IO introduces a pairwise formulation that combines delta-feature matching with anchor-based inference to transfer configuration-induced performance patterns learned at small scales to unseen target scales. Experimental results show that, when trained solely on measurements collected from 1-16 nodes, TSE-IO extrapolates the I/O time of IOR, S3D-IO, Flash-IO, FLASH, and LAMMPS to 128 nodes (2,048 processes), achieving a MAPE below 20% for the majority of evaluated workloads. After tuning, five representative applications exhibit significant I/O performance improvements, achieving speedups ranging from 16.38 × to 65.38 ×, and from 1.00 × to 2.79 × compared with expert user settings across different workloads, thereby demonstrating the effectiveness and generality of TSE-IO.
- 2 November 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
Open AccessTSE-IO: Two-Stage Extrapolative I/O Modeling and Tuning via Pairwise Learning
Collecting I/O data for parallel programs at large node scales is prohibitively expensive and has long been a major bottleneck for I/O performance modeling. Existing approaches typically depend on extensive training samples collected on large-scale systems, which incur high cost and generalize poorly across applications and environments. To address these challenges, we propose TSE-IO, a T wo- S tage framework for E xtrapolative I/O performance modeling and tuning. TSE-IO decomposes cross-scale performance prediction into two linked tasks: first, extrapolating application-level I/O behavior from small-scale measurements to the target scale. And second, predicting target-scale performance by jointly modeling the extrapolated behavior and tunable I/O-stack configurations. For the second stage, TSE-IO introduces a pairwise formulation that combines delta-feature matching with anchor-based inference to transfer configuration-induced performance patterns learned at small scales to unseen target scales. Experimental results show that, when trained solely on measurements collected from 1-16 nodes, TSE-IO extrapolates the I/O time of IOR, S3D-IO, Flash-IO, FLASH, and LAMMPS to 128 nodes (2,048 processes), achieving a MAPE below 20% for the majority of evaluated workloads. After tuning, five representative applications exhibit significant I/O performance improvements, achieving speedups ranging from 16.38 × to 65.38 ×, and from 1.00 × to 2.79 × compared with expert user settings across different workloads, thereby demonstrating the effectiveness and generality of TSE-IO.
- Open Access
- 25 June 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 23 (2) , 1-24
BLG-Tuning: Benchmark-Based Low-Cost General-Purpose I/O Modeling and Tuning
I/O performance has become a major bottleneck for many data-intensive applications. Each layer of the parallel I/O stack provides parameters that can optimize I/O performance, but determining the optimal performance parameters based on the operating configuration is a challenge. Previous work has required separate performance models for different programs for tuning, which is very costly in term of measurement data. We propose BLG-Tuning: a B enchmark-based L ow-cost G eneral-purpose I/O Modeling and Tuning. BLG-Tuning maps application I/O loads to benchmark parameters and uses the benchmark-trained performance model to achieve I/O performance prediction and thus avoid the additional computing and communication overhead for measurement. For applications, BLG-Tuning collects the application characteristics to calibrate the performance model and improve prediction accuracy. Experience shows that BLG-Tuning predicts the I/O time of MADbench2, Flash-IO, S3D-IO, BT-IO, and LAMMPS with MAPE of 22.2%, 18.2%, 29.3%, 21.5%, and 34.8%, respectively. After tuning, the five applications obtain I/O speedup from 5.6× to 27.3×.
- 25 June 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 23 (2) , 1-24
Open AccessBLG-Tuning: Benchmark-Based Low-Cost General-Purpose I/O Modeling and Tuning
I/O performance has become a major bottleneck for many data-intensive applications. Each layer of the parallel I/O stack provides parameters that can optimize I/O performance, but determining the optimal performance parameters based on the operating configuration is a challenge. Previous work has required separate performance models for different programs for tuning, which is very costly in term of measurement data. We propose BLG-Tuning: a B enchmark-based L ow-cost G eneral-purpose I/O Modeling and Tuning. BLG-Tuning maps application I/O loads to benchmark parameters and uses the benchmark-trained performance model to achieve I/O performance prediction and thus avoid the additional computing and communication overhead for measurement. For applications, BLG-Tuning collects the application characteristics to calibrate the performance model and improve prediction accuracy. Experience shows that BLG-Tuning predicts the I/O time of MADbench2, Flash-IO, S3D-IO, BT-IO, and LAMMPS with MAPE of 22.2%, 18.2%, 29.3%, 21.5%, and 34.8%, respectively. After tuning, the five applications obtain I/O speedup from 5.6× to 27.3×.
- Open Access
- 19 May 2026
- Published by Association for Computing Machinery (ACM)
- p. 271-279
A Portable GPU Kernel Performance Modeling Method Based on LLVM IR Dynamic Feature Prediction
- 19 May 2026
- Published by Association for Computing Machinery (ACM)
- p. 271-279
Open AccessA Portable GPU Kernel Performance Modeling Method Based on LLVM IR Dynamic Feature Prediction
- 12 August 2025
- Published by Elsevier in Future Generation Computer Systems
- Vol. 175
DP-SWAP: Fast Swapping Strategy Based on Dynamic Programming
- 12 August 2025
- Published by Elsevier in Future Generation Computer Systems
- Vol. 175
DP-SWAP: Fast Swapping Strategy Based on Dynamic Programming
- Open Access
- 20 March 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 22 (1) , 1-27
ATP: Achieving Throughput Peak for DNN Training via Smart GPU Memory Management
Due to the limited GPU memory, the performance of large DNNs training is constrained by the unscalable batch size. Existing studies partially address the issue of GPU memory limit through tensor recomputation and swapping, but overlook the exploration of optimal performance. In response, we propose ATP, a recomputation and swapping based GPU memory management framework that aims to maximize training performance by breaking GPU memory constraints. ATP utilizes a throughput model and we propose to evaluate the theoretical peak performance achievable by DNN training on GPU, and provide the optimum memory size required for recomputation and swapping. We optimize the mechanisms for GPU memory pool and CUDA stream control, employ an optimization method to search for specific tensors requiring recomputation and swapping, thereby bringing the actual DNN training performance on ATP closer to theoretical values. Evaluations with different types of large DNN models indicate that ATP achieve throughput improvements ranging from 1.14∼ 1.49×, while support model training exceeding the GPU memory limit by up to 9.2×.
- 20 March 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 22 (1) , 1-27
Open AccessATP: Achieving Throughput Peak for DNN Training via Smart GPU Memory Management
Due to the limited GPU memory, the performance of large DNNs training is constrained by the unscalable batch size. Existing studies partially address the issue of GPU memory limit through tensor recomputation and swapping, but overlook the exploration of optimal performance. In response, we propose ATP, a recomputation and swapping based GPU memory management framework that aims to maximize training performance by breaking GPU memory constraints. ATP utilizes a throughput model and we propose to evaluate the theoretical peak performance achievable by DNN training on GPU, and provide the optimum memory size required for recomputation and swapping. We optimize the mechanisms for GPU memory pool and CUDA stream control, employ an optimization method to search for specific tensors requiring recomputation and swapping, thereby bringing the actual DNN training performance on ATP closer to theoretical values. Evaluations with different types of large DNN models indicate that ATP achieve throughput improvements ranging from 1.14∼ 1.49×, while support model training exceeding the GPU memory limit by up to 9.2×.
- Open Access
- 14 September 2024
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 21 (3) , 1-25
An Example of Parallel Merkle Tree Traversal: Post-Quantum Leighton-Micali Signature on the GPU
The hash-based signature (HBS) is the most conservative and time-consuming among many post-quantum cryptography (PQC) algorithms. Two HBSs, LMS and XMSS, are the only PQC algorithms standardised by the National Institute of Standards and Technology (NIST) now. Existing HBSs are designed based on serial Merkle tree traversal, which is not conducive to taking full advantage of the computing power of parallel architectures such as CPUs and GPUs. We propose a parallel Merkle tree traversal (PMTT), which is tested by implementing LMS on the GPU. This is the first work accelerating LMS on the GPU, which performs well even with over 10,000 cores. Considering different scenarios of algorithmic parallelism and data parallelism, we implement corresponding variants for PMTT. The design of PMTT for algorithmic parallelism mainly considers the execution efficiency of a single task, while that for data parallelism starts with the full utilisation of GPU performance. In addition, we are the first to design a CPU-GPU collaborative processing solution for traversal algorithms to reduce the communication overhead between CPU and GPU. For algorithmic parallelism, our implementation is still 4.48× faster than the ideal time of the state-of-the-art traversal algorithm. For data parallelism, when the number of cores increases from 1 to 8,192, the parallel efficiency is 78.39%. In comparison, our LMS implementation outperforms most existing LMS and XMSS implementations.
- 14 September 2024
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 21 (3) , 1-25
Open AccessAn Example of Parallel Merkle Tree Traversal: Post-Quantum Leighton-Micali Signature on the GPU
The hash-based signature (HBS) is the most conservative and time-consuming among many post-quantum cryptography (PQC) algorithms. Two HBSs, LMS and XMSS, are the only PQC algorithms standardised by the National Institute of Standards and Technology (NIST) now. Existing HBSs are designed based on serial Merkle tree traversal, which is not conducive to taking full advantage of the computing power of parallel architectures such as CPUs and GPUs. We propose a parallel Merkle tree traversal (PMTT), which is tested by implementing LMS on the GPU. This is the first work accelerating LMS on the GPU, which performs well even with over 10,000 cores. Considering different scenarios of algorithmic parallelism and data parallelism, we implement corresponding variants for PMTT. The design of PMTT for algorithmic parallelism mainly considers the execution efficiency of a single task, while that for data parallelism starts with the full utilisation of GPU performance. In addition, we are the first to design a CPU-GPU collaborative processing solution for traversal algorithms to reduce the communication overhead between CPU and GPU. For algorithmic parallelism, our implementation is still 4.48× faster than the ideal time of the state-of-the-art traversal algorithm. For data parallelism, when the number of cores increases from 1 to 8,192, the parallel efficiency is 78.39%. In comparison, our LMS implementation outperforms most existing LMS and XMSS implementations.
- 11 September 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computers
- Vol. 74 (1) , 15-28
CUSPX: Efficient GPU Implementations of Post-Quantum Signature SPHINCS +
Quantum computers pose a serious threat to existing cryptographic systems. While Post-Quantum Cryptography (PQC) offers resilience against quantum attacks, its performance limitations often hinder widespread adoption. Among the three National Institute of Standards and Technology (NIST)-selected general-purpose PQC schemes, SPHINCS${}^{+}$ is particularly susceptible to these limitations. We introduce CUSPX ( CU DA SP HIN CS${}^{+}$ ), the first large-scale parallel implementation of SPHINCS${}^{+}$ capable of running across 10,000 cores. CUSPX leverages a novel three-level parallelism framework, applying it to algorithmic parallelism , data parallelism , and hybrid parallelism . Notably, CUSPX introduces parallel Merkle tree construction algorithms for arbitrary parallel scales and several load-balancing solutions, further enhancing performance. By treating tasks parallelism as the top level of parallelism, CUSPX provides a four-level parallel scheme that can run with any number of tasks. Evaluated on a single GeForce RTX 3090 using the SPHINCS${}^{+}$ -SHA-256-128s-simple parameter set, CUSPX achieves a single task's signature generation latency of 0.67 ms, demonstrating a 5,105$\times$ speedup over a single-thread version and an 18.50$\times$ speedup over the previous fastest implementation.- 11 September 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computers
- Vol. 74 (1) , 15-28
CUSPX: Efficient GPU Implementations of Post-Quantum Signature SPHINCS +
Quantum computers pose a serious threat to existing cryptographic systems. While Post-Quantum Cryptography (PQC) offers resilience against quantum attacks, its performance limitations often hinder widespread adoption. Among the three National Institute of Standards and Technology (NIST)-selected general-purpose PQC schemes, SPHINCS${}^{+}$ is particularly susceptible to these limitations. We introduce CUSPX ( CU DA SP HIN CS${}^{+}$ ), the first large-scale parallel implementation of SPHINCS${}^{+}$ capable of running across 10,000 cores. CUSPX leverages a novel three-level parallelism framework, applying it to algorithmic parallelism , data parallelism , and hybrid parallelism . Notably, CUSPX introduces parallel Merkle tree construction algorithms for arbitrary parallel scales and several load-balancing solutions, further enhancing performance. By treating tasks parallelism as the top level of parallelism, CUSPX provides a four-level parallel scheme that can run with any number of tasks. Evaluated on a single GeForce RTX 3090 using the SPHINCS${}^{+}$ -SHA-256-128s-simple parameter set, CUSPX achieves a single task's signature generation latency of 0.67 ms, demonstrating a 5,105$\times$ speedup over a single-thread version and an 18.50$\times$ speedup over the previous fastest implementation.- 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
- 1 November 2023
- Published by Elsevier in Future Generation Computer Systems
- Vol. 152, 160-169
pommDNN: Performance optimal GPU memory management for deep neural network training
- 1 November 2023
- Published by Elsevier in Future Generation Computer Systems
- Vol. 152, 160-169
pommDNN: Performance optimal GPU memory management for deep neural network training
- 22 September 2023
- Published by Springer Nature in The Journal of Supercomputing
- Vol. 80 (4) , 5042-5072
Parallel implementations of post-quantum leighton-Micali signature on multiple nodes
- 22 September 2023
- Published by Springer Nature in The Journal of Supercomputing
- Vol. 80 (4) , 5042-5072
Parallel implementations of post-quantum leighton-Micali signature on multiple nodes