Computer systems have many components. Designing a highly efficient algorithm requires carefully examining the intersection of the algorithm design space and modern computer architecture. The design of algorithms for simultaneous execution on multi-core CPUs and graphics processing units (GPUs) is becoming increasingly important, particularly as the world’s fastest supercomputers rely on GPUs to obtain high computational throughput. The GPU contains thousands of cores that can rapidly carry out computation and have very high on-card memory bandwidth. Working with Prof. Gowanlock, Zane Fink and Jordan Wright developed a parallel multiway merge algorithm, which is fundamental to the field of databases and other application areas. The team showed that significant performance gains can be achieved over CPU-only approaches by splitting the work between multiple CPU cores and the GPU. The team is currently studying other hybrid CPU/GPU algorithms. More information can be found in their paper cited below.
Michael Gowanlock, Ben Karsin, Zane Fink, and Jordan Wright. 2019. Accelerating the Unacceleratable: Hybrid CPU/GPU Algorithms for Memory-Bound Database Primitives. In Proceedings of the 15th International Workshop on Data Management on New Hardware (DaMoN’19). ACM, New York, NY, USA, Article 7, 11 pages. DOI: https://doi.org/10.1145/3329785.3329926