基于JIAJIA系统的消息传递和共享存储编程模式比较

为了研究基于软件DSM系统的OpenMP实现，该文以一类具有代表性的用户题为例，分别测试了其基于JIAJIA系统的两种实现方式的加速比：一种是用JIAJIA提供的消息传递系统调用，实现一个类MPI版本（方式1）；另一种是用多个处理机对共享数组的读写来替代消息传递系统调用（方式2）。测试结果发现，对少量处理机系统，两种方式还具有可比性，但是，随着处理机数的增多，共享存储应用的性能急剧下降。通过对测试结果的分析及对用户题的进一步测试，发现方式2的时间主要花费在做一致性处理和缺页中断处理而导致的大量小消息通信上。测试表明，JIAJIA共享存储程序一般会比MPI程序导致更重的网络负载。要在JIAJIA共享存储基础之上建立一种实用的共享并行计算环境，尤其在支持OpenMP等共享编程语言方面，还有待进一步工作。

Performance Comparison of Two Programming Patterns of JIAJIA: Message Passing and Shared Memory

In order to implement OpenMP on software DSM system, the paper evaluates the performance of the two programming patterns of JIAJIA:message passing and shared memory. In the first pattern, the jia_send and jia_recv are used to implement a MPI-like program. In the second pattern, shared data is used to implement a shared memory program. The results are found that the speedup of the second pattern is much lower than that of the first one, the lower the number of processors grows higher. From another evaluation, the result is got that a great number of short message for coherence protocal is the performance bottleneck of the second pattern. Therefore, JIAJIA must be perfected in order to support shared parallel computing and compiler such as OpenMP.