http://mail.nl.linux.org/linux-mm-www/2007-01/msg00000.html Dear Linux MM developers, I would like to introduce a new buffer cache management scheme. Existing buffer cache management scheme considers only temporal locality of the blocks in buffer cache and reduces number of buffer cache misses, while our scheme considers both temporal locality and sequentiality of the blocks and reduces the time costed by accessing disks. Due to the mechanical positioning components in hard drives, accesses to randomly placed blocks costs much more time than those to sequentially placed blocks (10ms vs. 0.1ms). Reducing ONE miss to random block saves more time than reducing TEN misses to sequential blocks. Our scheme improves system performance by giving high priority to random blocks to stay in the buffer cache so that the costly accesses to random blocks can be reduced or avoided. Please access a related technical paper published in Proc. of the 4th USENIX Conference on File and Storage Technologies(FAST'05) http://www.cse.ohio-state.edu/hpcs/WWW/HTML/publications/papers/TR-05-14.pdf We have implemented a prototype based on 2.6.11.9 kernel. The patch and tarball can be accessed at http://www.cse.ohio-state.edu/~dingxn/ Your comments and suggestions are more than welcome. Regards. Xiaoning Ding |