An Optimized RDMA Scheme for Efficient VM Synchronization in QEMU Checkpoints

In virtualization, ensuring high availability and reliability for critical business applications can be achieved through Fault Tolerance (FT) solutions. QEMU, a popular open-source virtualization platform, employs Coarse-grained lock-stepping (COLO) to maintain high availability by synchronizing states of the Primary Virtual Machine (PVM) and the Secondary Virtual Machine (SVM). The high-speed synchronization process of COLO checkpoints involves efficient data transfer, which is crucial for service continuity. In Ethernet-based virtual machine data transmission, Remote Direct Memory Access (RDMA) offers superior performance in reducing latency and increasing throughput compared to TCP/UDP. However, the original RDMA implementation in QEMU’s COLO process is limited, constraining its efficiency. This paper proposes an improved RDMA scheme for QEMU FT, based on the compressed transmission of dirty pages, enhancing COLO checkpoint synchronization in bandwidth-constrained environments. In experimental evaluations conducted in a 10 Gbps network environment, our scheme achieved a $5.26 \%$ to $14.09 \%$ improvement in Oracle tests, and a $\mathbf{1. 7 7 \%}$ to $\mathbf{3 1. 0 3 \%}$ improvement across various tests in MySQL. These results demonstrate the effectiveness of our scheme in enhancing FT performance under different database workloads, further validating its potential for widespread application in cloud data center virtualization environments.