Exploration of SMP-Aware DAO Memory Performance Issues-Final Report 2002
The performance of many LLNL applications is dominated by the cost of main memory accesses. Worse, many current trends in computer architecture will lead to substantial degradation of the percentage of peak performance obtained by these codes. This project yields novel techniques that alleviate this problem in SMP-based systems, which are common at LLNL. Further, our techniques will complement other emerging mechanisms for improving memory system performance, such as processor-in-memory. The exploration of existing dynamic access ordering (DAO) mechanisms adapted to SMPs and the development of new memory performance optimization techniques will lead to significant improvements in run times for LLNL applications on future computing platforms, effectively increasing the size of the platform. In this project, we have focused on a range of techniques to overcome the performance bottleneck of current multiprocessor systems and to increase the single-node efficiency. These efforts include the design and implementation of a toolset to analyze memory access patterns of applications, the exploration of regularity metrics and their use to classify code behavior, and a set of microbenchmarks to assess and quantify the performance of SMP memory systems. We will make these tools available to the general laboratory user community to help the evaluation and optimization of LLNL applications. In addition, we explored the use of Dynamic Access Ordering (DAO) techniques in the realm of shared memory multiprocessors. The most critical part of the latter is the need to maintain coherence among reordered accesses due to possible aliasing. We have worked on several design alternatives to guarantee consistency in such systems without changing the user environment. This guarantees that such novel memory systems will be directly applicable for existing and future HPC codes at LLNL.
- Research Organization:
- Lawrence Livermore National Lab. (LLNL), Livermore, CA (United States)
- Sponsoring Organization:
- US Department of Energy (US)
- DOE Contract Number:
- W-7405-ENG-48
- OSTI ID:
- 15003253
- Report Number(s):
- UCRL-ID-151790; TRN: US200422%%43
- Resource Relation:
- Other Information: PBD: 4 Feb 2003
- Country of Publication:
- United States
- Language:
- English
Similar Records
An Integrated Performance Visualizer for MPI/OpenMP Programs
Experience with mixed MPI/threaded programming models