Recently, there has been much research about cluster computing to achieve high performance using low cost nodes connected with high speed networks. In cluster computing, I/O operations are still a performance bottleneck because an I/O speed is much slower than any other component``s speed. To provide users with high performance I/O, many cluster I/O subsystems and mechanisms have been proposed such as cluster file systems, Cooperative Cache, collective I/O operations and so on.
In this work, we propose the new Cooperative Cache architecture and RW-phase protocol for scientific applications to improve the write performance of cluster file systems.
First, we propose the home-based Cooperative Cache. In the home-based Cooperative Cache, each block has it own home node and the home node is statically and globally determined. All information about a cached block is managed by the home node of the block and the home node does the write buffering as well as the read caching of the block. Using the home-based Cooperative Cache, a client can directly fetch a block from the home node of the block without contacting the manager as long as the block is in Cooperative Cache. Additionally, using the write buffering in home nodes, the home-based Cooperative Cache can improve write performance of I/O subsystems and can easily solve a false sharing problem in write-back Cooperative Cache because all small written data to a block is buffered in the home node of the block. For efficiently managing cached block in the home-based Cooperative Cache, we propose the new replacement policy named Pcs-LRU. Using Pcs-LRU, the home blocks can exist longer than the replication blocks in Cooperative Cache without the loss of the hit ratio. We also support the data migration for the making up the weak point of the static home management. Using the data migration, the data can be migrated to the frequently used node if the data is not accessed in the home node.
Second, to eliminate redundant o...