Lab for High Performance Computing SERC, Indian Institute of Science
Home | People | Research | Awards/Honours | Publications | Lab Resources | Gallery | Contact Info | Sponsored Research
Tech. Reports | Conferences / Journals | Theses / Project Reports

Hidden Costs in Avoiding False Sharing in Software DSMs

Proceedings of the International Conference on High Performance Computing (HiPC-01)
Hyderabad, India, December 2001


  1. K.V. Manjunath, Hewlett-Packard ISO, Bangalore; Department of Computer Science and Automation
  2. R. Govindarajan, Supercomputer Education and Research Centre; Department of Computer Science and Automation


Page based software DSMs experience high degrees of false sharing especially in irregular applications with fine grain sharing granularity. The overheads due to false sharing is considered to be the dominant factor limiting the performance of software DSMs. Several approaches have been proposed in the literature to reduce/eliminate false sharing. In this paper, we evaluate two of these approaches, viz., the Multiple Writer approach and Emulated Fine Grain Sharing (EmFiGS) approach. Our evaluation strategy is two pronged: firstly, we use a novel implementation independent analysis which uses overhead counts to compare the different approaches. The performance of EmFiGS approach is significantly worse, by a factor of 1.5 to as much as 90 times, compared to the Multiple Writer Approach. In many cases, EmFiGS approach performs worse that even a single writer lazy release protocol which experiences very high overheads due to false sharing. Our analysis shows that the benefits gained by eliminating false sharing are far outweighed by the performance penalty incurred due to the reduced exploitation of spatial locality in the EmFiGS approach. The implementation independent nature of our analysis implies that any implementation of the EmFiGS approach is likely to perform significantly worse than the Multiple Writer approach. Secondly, we use experimental evaluation to validate and complement our analysis. Experimental results match well with our analysis. Also the execution times of the application follow the same trend as in our analysis, reinforcing our conclusions.


Full Text