Multi-GPU Design and Performance Evaluation of Homomorphic Encryption on GPU Clusters

We present a multi-GPU design, implementation and performance evaluation of the Halevi-Polyakov-Shoup (HPS) variant Fan-Vercauteren (FV) levelled Fully Homomorphic Encryption (FHE) scheme. Our design follows data parallelism approach uses partitioning methods to distribute workload in FV primitives evenly across available GPUs. The is put address space runtime requirements FHE computations. It also suitable for distributed-memory architectures, includes efficient GPU-to-GPU exchange protocols. Moreover, it user-friendly as user intervention not required task decomposition, scheduling or load balancing. implement evaluate our on two homogeneous heterogeneous NVIDIA GPU clusters: K80, customized P100. provide comparison with recent shared-memory-based multi-core CPU using homomorphic circuits workloads: vector addition multiplication. we use Levelled-FHE inference circuit Convolutional Neural Networks (CNNs) perform homomorphically image classification encrypted images from MNIST CIFAR - 10 datasets. provides 1 3 orders magnitude speedup compared operations. In terms scalability, shows reasonable scalability curves when GPUs are fully connected.