NPB Performance Evaluation Of Tera-Scale Clusters And Implementation Of Parallel Non-Numerical Algorithm With Performance Analysis

We should pay attention to two facts in the rapid progress of high performance computing, one is that the peak performance of parallel computer is in fast progress and it has got the level of 100 Tflops, cluster with high performance/cost value has now become the main architecture and is adopted in more applications; At the same time, the sustained performance of parallel applications is very low compared with the peak performance of the computer, most parallel applications can only exploit below 20 percent of the peak performance.The real application performance is more important than peak performance, and it is what we care about most. The shortage of parallel application and low level of sustained performance has become the bottleneck in the progress of high performance computing. Both parallel computer and parallel applications affect the real performance, so we carried out application oriented performance benchmarking and application performance analysis on tera-scale cluster systems.NPB benchmarking was performed on three domestic tera-scale cluster systems with emphasis on the performance characteristics and trends when carrying out tera-scale parallel computing on systems with thousands of processors. The effects of different system configurations (processor, interconnection network, etc.) on final NPB performance were analyzed and it is found that the programs in NPB suites got their best performance on different clusters. Through further analysis, we found out that the scalability of NPB programs can reach hundreds of processors, but can't reach thousands of processors. Most of NPB programs can only exploit around 10% of system peak performance, the scalability of cluster systems and real application performance on tera-scale cluster systems need further improvement. For manufactures of tera-scale cluster systems with thousands of processors, theperformance of collective communication and fine-grained message passing needs further improvement.Performance research of parallel non-numerical applications is also very important. We developed a parallel data-mining program (association rule mining) and tested its speedup performance on two cluster systems. With good data partition and optimized data structures, this program has good parallel performance.The main works of my thesis are:· I performed NPB benchmarking on three domestic tera-scale cluster systems. Analyzed the effects of different system configurations on final NPB performance, Studied the sustained performance and scalability of NPB programs with thousands processors.· I developed a parallel data mining system (association rule mining) and tested its speedup on two cluster systems. Using the characteristics of this program analyzed the main factors that affected the performance.