Large-scale Bundle Adjustment (BA) is the key for many 3D vision applications (e.g., Structure-from-Motion and SLAM). Though important, large-scale BA is still poorly supported by existing BA libraries (e.g., Ceres and g2o). These libraries under-utilise accelerators (i.e., GPUs), and they lack algorithms to distribute BA computation constrained by the memory on a single device. In this paper, we propose MegBA, a high-performance and distributed library for large-scale BA. MegBA has a novel end-to-end vectorised BA algorithm that can fully exploit the massive parallel cores on GPUs, thus speeding up the entire BA computation. It also has a novel distributed BA algorithm that can automatically partition BA problems, and solve BA sub-problems using distributed GPUs. The GPUs synchronise intermediate solving state using network-efficient collective communication, and the synchronisation is designed to minimise communication cost. MegBA has a memory-efficient GPU runtime and exposes g2o-compatible APIs. Experiments show that MegBA can out-perform state-of-the-art BA libraries (i.e., Ceres and DeepLM) by up to 33x and 3.3x respectively, in public large-scale BA benchmarks. The code of MegBA is available at: \url{https://github.com/MegviiRobot/MegBA}.