Hashinator: a portable hybrid hashmap designed for heterogeneous high performance computing

Abstract

Scientific computing has become increasingly parallel and heterogeneous with the proliferation of graphics processing unit (GPU) use in data centers, allowing for thousands of simultaneous calculations accessing high-bandwidth memory. Adoption of these resources may require re-design of scientific software. Hashmaps are a widely used data structure linking unsorted unique keys with values for fast data retrieval and storage. Several parallel libraries exist for performing hashmap operations utilizing GPU hardware, but none have yet supported GPUs and CPUs interchangeably. We introduce Hashinator, a novel portable hashmap designed to operate efficiently on both CPUs and GPUs using CUDA or HIP/ROCm Unified Memory, offering host access methods, in-kernel access methods, and efficient GPU offloading capability on both NVIDIA and AMD hardware. Hashinator utilizes open addressing with Fibonacci hashing and power-of-two capacity. By comparing against existing implementations, we showcase the excellent performance and flexibility of Hashinator, making it easier to port scientific codes that rely heavily on the use of hashmaps to heterogeneous architectures.