Search

The sparse triangular solver (SpTRSV) is mostly used for scientific and engineering applications. The structured-grid triangular solver of regular dependencies (STRSV) is a special kind of SpTRSV. Some general SpTRSVs that disregards the regularity of the matrix are unsuitable for solving this problem. This paper proposes an efficient parallel algorithm for STRSV on the SW26010 (a kind of China independently designed many-core processors), namely swStructTRSV. The algorithm makes full use of the fine-grained and low latency communication characteristics of the SW26010 to reduce the waiting time for synchronization, maximizes the regularity of access to improve memory access bandwidth, and achieves overlap between memory access and computation simultaneously. Moreover, the idea of the algorithm can be extended to incomplete LU factorization (ILU factorization) because of consistent dependencies. The experimental results on a core group(8 * 8 network composed of 64 cores) of SW26010 show that swStructTRSV can achieve an average speedup of over 30 in the sequential version. swStructTRSV on SW26010 achieves solving speedups of 2.2 and 6.3 over the fast STRSV (fSpTRSV) previously implemented on SW26010 and MKL on Intel Xeon Gold 6132, respectively. swStructTRSV significantly outperforms cuSparse on NVIDIA TITAN RTX in terms of overall execution time. [ABSTRACT FROM AUTHOR]

This paper presents a massively parallel approach of the multilevel fast multipole algorithm (PMLFMA) on homegrown many-core SW26010 cluster of China, noted as (SW-PMLFMA), for 3-D electromagnetic scattering problems. In this approach, the multilevel fast multipole algorithm (MLFMA) octree is first partitioned among management processing elements (MPEs) of SW26010 processors following the ternary partitioning scheme using the message passing interface (MPI). Then, the computationally intensive parts of the PMLFMA on each MPI process, matrix filling, aggregation and disaggregation are accelerated by using all the 64 computing processing elements (CPEs) in the same core group of the MPE via the Athread parallel programming model. Different parallelization strategies are designed for many-core accelerators to ensure a high computational throughput. In coincidence with the special characteristic of local Lagrange interpolation, the compressed sparse row (CSR) and the compressed sparse column (CSC) sparse matrix storage format is used for storing interpolation and anterpolation matrices, respectively, together with a specially designed cache mechanism of hybrid dynamic and static buffers using the scratchpad memory (SPM) to improve data access efficiency. Numerical results are included to demonstrate the efficiency and versatility of the proposed method. The proposed parallel scheme is shown to have excellent speedup. [ABSTRACT FROM AUTHOR]

As the size of graph data grows exponentially, it is usually processed by parallel and distributed computing environment. During parallel processing, the first step is to divide the graph data into many subgraphs and then place them on different computational nodes. However, the existing graph clustering methods cannot adapt to parallel processing or improve computing performance. The main reason is that these methods only focus on the graph data but do not care about the graph's application algorithm. Therefore, this paper addresses a graph clustering method that supports parallel processing by perceiving the application algorithm semantics. We specifically focus on the characteristic of "tight inside and loose outside". Based on this characteristic, we give a new formula for calculating the modularity of the subgraph. In a graph application, graph data and its associated application algorithm are not separated from each other. Thus, we also focus on the execution patterns of the application algorithm. Combining the modularity of a subgraph and execution patterns of an application algorithm, we present the critical concept of semantic serial degree as a new criterion for big graph clustering. Consequently, we propose a graph clustering method that effectively balances the importance between the graph data and its associated application algorithm. By achieving this balance, our approach ensures that the final clustering results are more suitable for parallel and distributed processing. Extended experiments show that our proposed graph clustering method is more general and compatible with traditional clustering methods. Compared with FastUnfolding, a state-of-the-art graph clustering method, the completion times of an application algorithm are significantly reduced due to the help of the proposed graph clustering method. [ABSTRACT FROM AUTHOR]