~/talk

GPUKV: Towards a GPU-Driven Computing on Key-Value SSD

Talkat 12th USENIX Workshop on Hot Topics in Storage and File Systems (HotStorage '20) · 2020-07-13..07-14

speakers

name:Min-Gyo Jeong

links:none on file

name:Chang-Gyu Lee

links:none on file

name:Donggyu Park

links:none on file

name:Sungyong Park

links:none on file

name:Youngjae Kim

links:none on file

name:Jungki Noh

links:none on file

name:Woosuk Chung

links:none on file

name:Kyoung Park

links:none on file

resources

video:https://youtu.be/c6y9QJI9bZ8

slides:none on file

summary

Describes a system that pushes key-value processing onto GPUs co-located with or attached to key-value SSDs, implementing KV-SSD access paths, CUDA kernels for hashing/lookup/compaction, and a custom SSD driver to move data without CPU copies. Builds a prototype combining an SK hynix KV-SSD and an NVIDIA GPU, measures higher throughput, lower end-to-end latency, and drastically reduced host CPU utilization compared with conventional CPU-only KVS paths.

AI-written summary · about