基于SIMD机器的优化数据传输的并行循环分割

本文提出一个基于分布式局存的ＳＩＭＤ机器的循环分割理论体系以优化运算中所需要的数据传输。该体系使用矩阵表示迭代空间、数据空间和数组存取式。我们引入数据传输概念，并建立一个简单有效的数据传输模型来评估数据在全局内存和局部内存之间的传输开销。最后，对于给定的循环嵌套，我们给出一个循环分割算法以获得优化循环块，使得循环嵌套中所需要的数据传输开销最小，并且大大减少了数据传输和计算的同步开销。实验结果证明了

PARALLEL LOOP PARTITIONING OPTIMIZED FOR DATA TRANSFER ON SIMD MACHINES

This paper presents a theoretical framework for automatically partitioning parallel loops to optimize the data movement on SIMD machines with distributed internal memory and global external memory. The framework uses matrix to represent iteration mapping, data mapping and array references. A system model is presented to characterize the SIMD machine. First, the data movement functions supplied by run-time system are analyzed to provide a good hasis for studying automatic loop and data partitioning. Then we introduce the notion of data footprints and build a simple but accurate data movement model to evaluate the overhead of the data movement between global external memory and distributed internal memory for the footprint. Based on the data movement model . this paper presents an algorithm to derive an optimal loop partitioning to minimize the number of data movement in the loop nest and optimize the synchronization between computation and data movement. These ideas are adopted in the design and implementation of a parallelizing C compiler targeted on the SIMD machine. The experimental results demonstrate that the optimal loop and data partitioning greatly reduce the overhead of data movement and improve the performance of many programs.