We have employed the Linux on-demand read-ahead framework in our campus wide storage service system call MeePo. The appropriate read-ahead mechanism can significantly increase IO throughput and improve user experiences by hiding network latency which is critical for real-time applications. Our strategy is based on the data accessing characteristic of MeePo system. The read-ahead framework uses the strategy profile which is generated according to the analysis of access trace of a typical user in a storage community. Our test deployment environment involves more than 5000 registered users as well as 150+ communities. Based on our observation that most of files in our system have either sequential or interleaved accessing patterns. In such scenario, client IO throughout could increase 12% for sequential stream and more than 180% improvement for interleaved stream.
展开▼
