Performance enhancement of fault-tolerant software DSM with message logging

Abstract

The probability of multiple failures in Software Distributed Shared Memory (SDSM) increases as the system size grows. Recently, the most popular technique of Fault-Tolerant Software Distributed Shared Memory (FT-SDSM) is to store the messages exchanged between communicating nodes in the proper storages with independent checkpointing, called message logging. With its popularity, however, the logging overheads are non-negligible during the failure-free execution, and hence in a few years, an impressive amount of researches have been focused on reducing such a non-negligible overhead. We have implemented a lightweight logging scheme on home-based SDSM. That is called remote logging where the remote memories are used for the logged data. With its lightweight advantage, however, the logged data in home nodes and back-up nodes are useless during the failure-free execution. We propose new log usable schemes that enable FT-SDSM with remote logging to be enhanced. All logged data in home nodes and back-up nodes are used to reduce the times stalled for updates of the invalid pages and to minimize the normal execution time. On a page fault, the invalid page can be updated by using the logged data instead of fetching a whole page from its home node. We have performed the experiments on eight PC clusters. We actually implemented our proposed schemes on FT-SDSM with remote logging. The experimental results show that our proposed log usable schemes outperform our based FT-SDSM which does not make use of any logs in some applications. When we apply our proposed log usable schemes in all, we reduce the number of messages by about 5 - 12 % and minimize the message amounts by about 11 - 78 %, and consequently total execution time becomes less about 13 % than our based model at the best case.