Full text
| |
| Author(s): |
Guilherme Guaglianoni Piccoli
Total Authors: 1
|
| Document type: | Master's Dissertation |
| Press: | Campinas, SP. |
| Institution: | Universidade Estadual de Campinas (UNICAMP). Instituto de Computação |
| Defense date: | 2016-04-18 |
| Examining board members: |
Edson Borin;
Philippe Olivier Alexandre Navaux;
Lucas Francisco Wanner
|
| Advisor: | Fernando Magno Quintão Pereira; Edson Borin |
| Abstract | |
Multicore processors are increasingly common, especially due to technological barrier imposed by physical limitations in the growth of processors clock frequency. Non-uniform memory architecture (NUMA) are spreading as a solution to the scalability of high performance computing applications - such architectures, however, still exhibits problems, especially in the case of memory accesses. Bad data placement can cause contention on memory access leading to significant decrease in applications performance. We present Selective Page Migration (SPM), a compiler optimization that, by analyzing loops and their properties along with arrays accessed within such loops, can perform selective page migration according to a heuristic. The goal is to mitigate contention issues and bad data placement in NUMA architectures with no need to modify source code or using specific hardware or operating systems. To achieve this goal, compilation transformations have been implemented so we can perform source code instrumentation to find data structures that, once migrated, lead to an increase in program performance. Also, a heuristic has been developed so we're able to assess whether the migration of those data structures are likely to become profitable or will potentially impair the application performance. We've achieved good results, with speedup of more than 5x for some benchmarks. We have compared SPM with another two mechanisms used with the same goal, and also presented a theoretical evaluation of a variety of techniques with the same end of SPM (AU) | |
| FAPESP's process: | 13/18794-3 - Compilation techniques to optimize the memory subsystem access |
| Grantee: | Guilherme Guaglianoni Piccoli |
| Support Opportunities: | Scholarships in Brazil - Master |