Write-Optimized and Consistent RDMA-based Non-Volatile Main Memory Systems
Xinxin Liu, Yu Hua, Xuan Li, Qifan Liu · 2021
To deliver high performance in cloud computing, many efforts leverage RDMA (Remote Direct Memory Access) in networking and NVMM (Non-Volatile Main Memory) in end systems. Due to no CPU involvement, one-sided RDMA becomes efficient to access the remote memory, and NVMM technologies have the strengths of non-volatility, byte-addressability and DRAM-like latency. However, due to the need to guarantee Remote Data Atomicity (RDA), the synergized scheme has to consume extra network round-trips, remote CPU participation and double NVMM writes. In order to address these problems, we propose a write-optimized log-structured NVMM design for Efficient Remote Data Atomicity, called Erda. In Erda, clients directly transfer data to the destination memory addresses in the logs on servers via one-sided RDMA writes without redundant copies and remote CPU consumption. To detect the atomicity of the fetched data, we verify a checksum without client-server coordination. We further ensure metadata consistency by leveraging an 8-byte atomic update in a hash table, which also contains the addresses of previous versions of data in the log for consistency. When a failure occurs, the server properly and efficiently restores to become consistent. Experimental results show that compared with state-of-the-art schemes, Erda reduces NVMM writes approximately by 50%, significantly improves throughput and decreases latency.