Design and Implementation of GPU Accelerated Active Storage in FastDFS
Zaihong He, Jishun Kuang, Yanjie Tan, Wenjie Liu, Bin Sheng · 2019
With the rapid increase of the data, the network bandwidth of the distributed storage system should be full utilized. Existing distributed storage models have higher network delay which leads to less efficiency in data-intensive services. Therefore, we propose a GPU-accelerated Active Storage (GAS) model for distributed file systems and implement it on FastDFS. GAS decreases the data transfer by executing the data clustering and filtering operations on storage nodes, and promotes the computing capacity of the storage server by calling GPU. For ensuring the active storage service commands be transmitted correctly, a communication protocol, called active storage communication protocol (ASCP) is designed. The experimental results show that, compared with traditional storage (TS) and active storage (AS), GAS reduces the execution time of SUM by up to 73.9% and 2.7%, lower the run time of AES by up to 91.9% and 70.3%, and shortens the time consumed of Mean-filter to 59.5% and 59.0% respectively.