Zhou Jin

Prof. Zhou Jin

Super Scientific Software Laboratory, China University of Petroleum, Beijing, China
orcid

Research Focus & Metrics

Academic Career Length

22Years (2005-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
1 May 2021
Verified
Conference Paper
Closed access
YNYuyao Niu
Yuyao Niu
ZLZhengyang Lu
Zhengyang Lu
MDMeichen Dong
Meichen Dong
Zhou JinZhou Jin
Zhou Jin
WLWeifeng Liu
Weifeng Liu
GTGuangming Tan
Guangming Tan
With the extensive use of GPUs in modern supercomputers, accelerating sparse matrix-vector multiplication (SpMV) on GPUs received much attention in the last couple of decades. A number of techniques, such as increasing utilization of wide vector units, reducing load imbalance and selecting the best formats, have been developed. However, the 2D spatial sparsity structure has not been well exploited in the existing work for SpMV on GPUs. In this paper, we propose an efficient tiled algorithm called TileSpMV for optimizing SpMV on GPUs through exploiting 2D spatial structure of sparse matrices. We first implement seven warp-level SpMV methods for calculating sparse tiles stored in a variety of formats, and then design a selection method to find the best format and SpMV implementation for each tile. We also adaptively extract nonzeros in the very sparse tiles into a separate matrix to maximize the overall performance. The experimental results show that our method is faster than state-of-the-art SpMV methods such as Merge-SpMV, CSR5 and BSR in most matrices of the full SuiteSparse Matrix Collection and delivers up to 2.61x, 3.96x and 426.59x speedups, respectively.

28 March 2022
Verified
Conference Paper
Open access
YNYuyao Niu
Yuyao Niu
ZLZhengyang Lu
Zhengyang Lu
HJHaonan Ji
Haonan Ji
SSShuhui Song
Shuhui Song
Zhou JinZhou Jin
Zhou Jin
WLWeifeng Liu
Weifeng Liu
Sparse general matrix-matrix multiplication (SpGEMM) is one of the most fundamental building blocks in sparse linear solvers, graph processing frameworks and machine learning applications. The existing parallel approaches for shared memory SpGEMM mostly use the row-row style with possibly good parallelism. However, because of the irregularity in sparsity structures, the existing row-row methods often suffer from three problems: (1) load imbalance, (2) high global space complexity and unsatisfactory data locality, and (3) sparse accumulator selection. We in this paper propose a tiled parallel SpGEMM algorithm named TileSpGEMM. Our algorithm sparsifies the tiled method in dense general matrix-matrix multiplication (GEMM), and saves each non-empty tile in a sparse form. Its first advantage is that the basic working unit is now a fixed-size sparse tile containing a small number of nonzeros, but not a row possibly very long. Thus the load imbalance issue can be naturally alleviated. Secondly, the temporary space needed for each tile is small and can always be in on-chip scratchpad memory. Thus there is no need to allocate an off-chip space for a large amount of intermediate products, and the data locality can be much better. Thirdly, because the computations are restricted within a single tile, it is relatively easier to select a fast sparse accumulator for a sparse tile. Our experimental results on two newest NVIDIA GPUs show that our TileSpGEMM outperforms four state-of-the-art SpGEMM methods cuSPARSE, bhSPARSE, NSPARSE and spECK in 139, 138, 127 and 94 out of all 142 square matrices executing no less than one billion flops for an SpGEMM operation, and delivers up to 2.78x, 145.35x, 97.86x and 3.70x speedups, respectively.

11 November 2023
Verified
Conference Paper
Open access
XFXu Fu
Xu Fu
BZBingbin Zhang
Bingbin Zhang
TWTengcheng Wang
Tengcheng Wang
WLWenhao Li
Wenhao Li
YLYuechen Lu
Yuechen Lu
WLWeifeng Liu
Weifeng Liu
Sparse direct solvers play a vital role in large-scale high performance computing in science and engineering. Existing distributed sparse direct methods employ multifrontal/supernodal patterns to aggregate columns of nearly identical forms and to exploit dense basic linear algebra subprograms (BLAS) for computation. However, such a data layout may bring more unevenness when the structure of the input matrix is not ideal, and using dense BLAS may waste many floating-point operations on zero fill-ins. In this paper, we propose a new sparse direct solver called PanguLU. Unlike the multifrontal/supernodal layout, our work relies on simpler regular 2D blocking and stores the blocks in their sparse forms to avoid any extra fill-ins. Based on the sparse patterns of the blocks, a variety of block-wise sparse BLAS methods are developed and selected for higher efficiency on local GPUs. To make PanguLU more scalable, we also adjust the mapping of blocks to processes for overall more balanced workload, and propose a synchronisation-free communication strategy considering the dependencies among different sub-tasks to reduce overall latency overhead. Experiments on two distributed heterogeneous platforms consisting of 128 NVIDIA A100 GPUs and 128 AMD MI50 GPUs demonstrate that PanguLU achieves up to 11.70x and 17.97x speedups over the latest SuperLU_DIST, and scales up to 47.51x and 74.84x on the 128 A100 and MI50 GPUs over a single GPU, respectively.

5 December 2021
Verified
Conference Paper
Closed access
JZJianqi Zhao
Jianqi Zhao
YWYao Wen
Yao Wen
YLYuchen Luo
Yuchen Luo
Zhou JinZhou Jin
Zhou Jin
WLWeifeng Liu
Weifeng Liu
ZZZhenya Zhou
Zhenya Zhou
Sparse LU factorization is one of the key building blocks of sparse direct solvers and often dominates the computing time of circuit simulation programs. Existing GPU-accelerated sparse LU factorization methods either offload relatively small dense matrix-matrix multiplications to GPU cores, or extract level-set information to parallelize elimination operations in each level. However, because of the insufficient parallelism, neither of the methods can saturate a large amount of compute units on modern GPUs.We in this paper propose a synchronization-free sparse LU factorization algorithm called SFLU. To saturate GPU cores, our method lets each thread block eliminate a column and runs all the thread blocks at the same time. Through communicating dependency information stored on global memory, all the thread blocks either busy wait to run or get updated by their previous columns. Because elimination of all the columns work concurrently, our method avoids any barrier synchronization and saturates GPU resources. By benchmarking over 1000 sparse matrices on an NVIDIA Titan RTX GPU, our SFLU outperforms SuperLU and GLU by a factor of on average 155.71 and 8.21 (up to 3585.62 and 252.66), respectively.

2 March 2023 in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
Verified
Journal Article
Closed access
DNDan Niu
Dan Niu
YDYichao Dong
Yichao Dong
Zhou JinZhou Jin
Zhou Jin
CZChuan Zhang
Chuan Zhang
QLQi Li
Qi Li
CSChangyin Sun
Changyin Sun
DC analysis is essential and still quite challenging in large-scale nonlinear circuit simulation. Pseudo transient analysis (PTA) is a widely-used and has great potential solver in the industry. However, the PTA convergence and simulation efficiency is still seriously affected by its stepping policy. This paper proposes an online stochastic stepping policy (OSSP) for PTA based on deep reinforcement learning (DRL). To achieve better policy evaluation and stronger stepping exploration ability, the dual soft Actor-Critic agents work with the proposed valuation splitting and online momental scaling, enabling our OSSP to intelligently encode PTA iteration status and online further adjust forward and backward time-step size for unseen test circuits without human intervention and domain knowledge, trained solely by RL from self-search. Our public sample buffer and priority sampling are also introduced to overcome the sparsity and imbalance of sample data. Numerical examples demonstrate that the proposed OSSP achieves a significant efficiency speedup (up to 47.0X less NR iterations) and convergence enhancement on unseen test circuits compared with the previous iter-based and SER-based stepping methods, in just one stepping iteration.

17 January 2022
Verified
Conference Paper
Closed access
YCYufei Chen
Yufei Chen
HPHaojie Pei
Haojie Pei
XDXiao Dong
Xiao Dong
Zhou JinZhou Jin
Zhou Jin
CZCheng Zhuo
Cheng Zhuo
Relentless semiconductor scaling and ever increasing device integration have resulted in the exponentially growing size of the back-end design, which makes back-end simulation very time- and resource-consuming. With the success in the computer vision community, deep learning seems a promising alternative to assist the back-end simulation. However, unlike computer vision tasks, most back-end simulation problems are mathematically and physically well-defined, e.g., power delivery network sign off and post-layout circuit simulation. It then brings broad interests in the community where and how to deploy deep learning in the back-end simulation flows. This paper discusses a few challenges that the deployment of deep learning models in back-end simulation have to confront and the corresponding opportunities for future research.