文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > Scalable Parallel Sparse Factorization with Left-Right Looking Strategy on Shared Memory Mu
侵权投诉

Scalable Parallel Sparse Factorization with Left-Right Looking Strategy on Shared Memory Mu

An efficient sparse LU factorization algorithm on popular shared memory multiprocessors is presented. Interprocess communication is critically important on these architectures- the algorithm introduces O(n) synchronization events only. No global barrier is

2

Olaf Schenk et al.

over a wide range of numerical properties and therefore the parallel sparse direct solver PARDISO has been integrated into complex semiconductor device and process simulation packages 4, 5]. The rapid and widespread acceptance of shared memory multiprocessor architectures, from the desktop to the high-end server, has now created a demand for e cient parallel sparse linear solvers on such shared memory architectures. The synchronization on modern multiprocessor architectures is one form of interprocess communication and can have large impact on program performance. Therefore, in section 2 we describe a method using O(n) synchronization events. The experimental results are presented in Section 3. Section 4 provides a short comparison of iterative and sparse direct methods for several semiconductor device simulation examples on di erent architectures. Section 5 contains a summary of our results.

There are three levels of parallelism that can be exploited by parallel direct methods. The rst one, elimination tree parallelism 8] is generally exploited in all parallel sparse direct packages. In practice, however, a large percentage of the computation occurs at the top levels of the tree and the sequential part of the elimination tree usually limits the e ciency. It is therefore important to consider more levels of parallelism. Another type of parallelism, called node level parallelism, solves in a hybrid approach at the top level of the tree the factorization of each supernode in parallel 9]. However, the size of the supernodes and the numbers of processors determine the e ciency. The third type, pipelining parallelism, instead, is suitable for a larger number of processors. In this case, a processor can start with the external factorization of all supernodes that are already factorized. Although a pipelining approach is di cult to realize in sparse direct solver packages, the pipelining parallelism is essential to achieve higher concurrency. Our parallel formulation of the sparse numerical factorization is based on the general framework described in 2, 10]. It successively computes fractions of supernodes called panels, which are introduced to increase load balancing and pipelining parallelism i

n the sequential part of the elimination tree. The algorithm with reduced synchronization is shown in Figure 1. First, the scheduler is initialized with all leaves of the elimination tree (line 2). Then, a group of p processes is created. Each process asks the scheduler for a new task until all panels have been factorized. The assignment of panel tasks (line 6) is completely dynamic, resulting in a signi cant improvement of load-balancing| if su ciently small tasks can be assigned last. That is the reason for the introduction of the panels. We have found from experimental results that the following strategy results in panel sizes that a large enough for Level 3 BLAS and small enough to exploit load balancing and pipelining parallelism: First, if (1< nproc 4) then each panel contains not more than 96 columns of the factor, if (4< nproc) then the maximum panel size is reduced to 72 columns. The scheduling is dynamic but not completely independent of other processes.

2 Parallel left-right looking strategy and asynchronous computation scheduling.

上一页第3页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email