Fast and Consistent Remote Direct Access to Non-volatile Memory

For performance benefits, recent trends of modern data centers tend to use NVM (Non-volatile Memory) as storage and utilize one-sided primitives of RDMA (Remote Direct Memory Access) to directly access NVM for I/O requests. However, one-sided RDMA doesn’t have durability semantics currently, and data may partially exist in the NVM when crashes happen, thus causing unexpected data inconsistency.

Existing solutions sacrifice either read or write performance in exchange for consistency guarantees. In this paper, we introduce a multi-version log-structuring design, named eFactory, which delivers high performance for both read and write, yet also provides data consistency. In eFactory, a multi-version list is built for each object to solve the problem of crash consistency, especially when multiple clients update the same object concurrently. To transfer data directly with RDMA write as well as reduce CRC overhead on the read critical path, a single background thread is assigned to conduct integrity verification and data persisting. Furthermore, eFactory proposes a hybrid read scheme to gain high performance for reads without losing consistency guarantee. With a durability flag embedded in the object, the client can detect the durability status of data and re-read it if necessary. Evaluations show that eFactory outperforms IMM and SAW (solutions that sacrifice write performance) by 0.42x-2.79x and 0.66x-2.85x for write respectively, while maintains comparable read performance with them. In addition, the read throughput of eFactory is 1.3x-1.96x and 1.24x-1.67x of Erda and Forca (solutions that guarantee consistency at the cost of reading performance) respectively.