| In order to fill the speed gap between processor and disk, many computer systemsuse buffer cache to buffer the disk data that has been accessed. At present, most of thebuffer management mechanism only use one kind buffer replacement policy, like LRU,MRU and so on. But seen from more and more applications, the characters of dataaccess are diversiform.Obviously one policy can not suit multiple applications.Therefore,unifying two policy-replacement-integrated and application-specificbuffer management mechanism, this thesis propose a new application-specific buffermanagement mechanism,which not only can support applications designate their owndata reference patterns and choose replacement policy according to reference patterndynamically, but also can detect correctness of reference pattern,which could beconsidered as a refer for application to adapt its actual reference patternThrough the analysis of the data access character from some applications,this thesisput forward four reference patterns: sequential reference pattem, looping referencepattem, probabilistic reference pattem and temporally-clustered reference pattern.According to these patterns, the system integrates three policies LRU, MRU and LFU,and designates the correlation between reference pattern and policy;To supply theapplication with feedback information, the system provides check-up function, to checkwhether the application's real reference pattern match the designated pattern;Afterunderstanding buffer management mechanism of Linux, we build our new mechanismbased on it.Finally, this thesis use trace-driven method to simulate the new mechanism ondisksim simulator. The simulation results show that, the disk I/O system with newmechanism can reduce the response latency and the disk I/O time, and has a betterperformance than the traditional system that use unique replacement policy,and it showsits advantage a certain extent comparing with ones using other application-specificmechanisms. |
PDF Full Text Request