A Dynamic Hash Table for the GPU
Electrical and Computer Engineering, University of California, Davis
arXiv:1710.11246 [cs.DC], (30 Oct 2017)
@article{ashkiani2017dynamic,
title={A Dynamic Hash Table for the GPU},
author={Ashkiani, Saman and Farach-Colton, Martin and Owens, John D.},
year={2017},
month={oct},
archivePrefix={"arXiv"},
primaryClass={cs.DC}
}
We design and implement a fully concurrent dynamic hash table for GPUs with comparable performance to the state of the art static hash tables. We propose a warp-cooperative work sharing strategy that reduces branch divergence and provides an efficient alternative to the traditional way of per-thread (or per-warp) work assignment and processing. By using this strategy, we build a dynamic non-blocking concurrent linked list, the slab list, that supports asynchronous, concurrent updates (insertions and deletions) as well as search queries. We use the slab list to implement a dynamic hash table with chaining (the slab hash). On an NVIDIA Tesla K40c GPU, the slab hash performs updates with up to 512 M updates/s and processes search queries with up to 937 M queries/s. We also design a warp-synchronous dynamic memory allocator, SlabAlloc, that suits the high performance needs of the slab hash. SlabAlloc dynamically allocates memory at a rate of 600 M allocations/s, which is up to 37x faster than alternative methods in similar scenarios.
November 5, 2017 by hgpu