The 21st USENIX Conference on File and Storage Technologies (FAST) was held in California, USA, which stands for the highest international research level in the field of storage systems. Prof. Yu Hua's team presented a paper titled ROLEX: A Scalable RDMA-oriented Learned Key-Value Store for Disaggregated Memory Systems, with Huazhong University of Science and Technology (HUST) as the only affiliation, and won the Best Paper Award, which is also the first team from China who has won this award.
Figure 1. The Best Paper Award Certificate at FAST 2023
This paper proposed a learned key-value scheme for one-sided RDMA operations in disaggregated memory systems, called ROLEX. By separating data modification and model retraining operations, the data accessed to the old model do not need to be retrained immediately, which significantly reduces the frequency of model retraining and system overheads. Based on the retraining separation model, the compute nodes directly access and modify remote data through one-sided RDMA operations. In order to reduce the consumption of network bandwidth, ROLEX uses an asynchronous and in-place retraining scheme to retrain the modified data on memory nodes with a small amount of computing resources. The performance evaluation leverages multiple real-world workloads. The evaluation results demonstrate that ROLEX can significantly improve system performance compared with state-of-the-art distributed indexing schemes.
Figure 2. The architecture of ROLEX
Figure 3. Prof.Yu Hua’s research team
FAST is a class-A international conference recommended by CCF, and for more than 20 years, has promoted the rapid development of many storage technologies, such as data deduplication, storage coding technology, non-volatile memory system and key-value storage. This conference represents the highest international level in the field of computer storage, and has extensive and far-reaching influence on the international academic and industrial communities.
