BTCache: A High Performance Concurrent Read Framework Based on Bit-Torrent Algorithm

During the computation stage of parallel seismic applications, it is usually a common need for a large amount of processes to concurrently read a large seismic dataset file from Cluster File Systems (CFS). When there are a great many of processes, for example, more than 100 processes, all processes reading from CFS will impose a huge pressure on the I/O subsystem of CFS. In this circumstance, the aggregated read bandwidth will decrease, which leads to that each process must spend a long time waiting for data to enter the computation stage. Thus, the efficiency of parallel seismic applications is affected severely. To solve this problem, we propose a concurrent read framework based on bit-torrent algorithm called BTCache. In BTCache, the whole dataset is divided into data blocks logically. Data blocks that are read by each process from CFS will be cached by a multi-level LRU-Cache constructed by user buffer, memory, and local disks on each computing node. Then, the bittorrent algorithm is used to transfer data among processes, so the pressure imposed on CFS will be reduced and the aggregated read bandwidth can be improved. Experimental results on a large high-performance computing cluster using 150 computing nodes show that BTCache can improve read bandwidth by 3.64 times compared with the concurrent read method. (Abstract)