일 | 월 | 화 | 수 | 목 | 금 | 토 |
---|---|---|---|---|---|---|
1 | 2 | 3 | ||||
4 | 5 | 6 | 7 | 8 | 9 | 10 |
11 | 12 | 13 | 14 | 15 | 16 | 17 |
18 | 19 | 20 | 21 | 22 | 23 | 24 |
25 | 26 | 27 | 28 | 29 | 30 | 31 |
- ssd
- rocksdb
- framework
- Machine Learning
- core dumped
- linux
- deep learning
- software
- USENIX
- memory
- Samsung
- FTL
- Git
- 키워드
- 시스템 소프트웨어
- kernel
- 시스템 프로그래밍
- overflow
- performance
- github
- 커널 프로그래밍
- Cache
- Flash Memory
- 포트 번호 변경
- Intel
- hardware
- Operating System
- storage system
- Today
- Total
Happy to visit my research note ^^
(관심 논문) Revitalizing the Forgotten On-Chip DMA to Expedite Data Movement in NVM-based Storage Systems 본문
(관심 논문) Revitalizing the Forgotten On-Chip DMA to Expedite Data Movement in NVM-based Storage Systems
Liam Lim 2023. 3. 21. 19:43Jingbo Su, Jiahao Li, and Luofan Chen, University of Science and Technology of China; Cheng Li, University of Science and Technology of China and Anhui Province Key Laboratory of High Performance Computing; Kai Zhang and Liang Yang, SmartX; Sam H. Noh, UNIST & Virginia Tech; Yinlong Xu, University of Science and Technology of China and Anhui Province Key Laboratory of High Performance Computing
February 21–23, 2023 • Santa Clara, CA, USA
USENIX Association
21st USENIX Conference on File and Storage Technologies
Abstract
NVM-based storage systems에서 실행되는 data-intensive applications는 DRAM과 NVM 간 data 이동 시 serious bottlenecks를 경험한다. 본 논문에서는 오래 전부터 존재했지만 최근 간과된 on-chip DMA의 사용을 권장하여 data movement를 가속화하는 세 가지 contribution (기여)을 제시한다.
- 첫째, comprehensive DMA study를 기반으로 latency-oriented optimization directions을 탐구하여 high-performance DMA module을 설계함으로써 I/O size threshold를 크게 낮추어 이점을 관찰할 수 있다.
- 둘째, DMA's limitations를 보상하고 전체적인 이점을 극대화하기 위해 적절한 scheduling과 load splitting을 통해 CPU와 DMA를 조율하는 새로운 data movement engine인 Fastmove를 제안한다.
- 마지막으로, general kernel-based design, simple APIs, and DMX file system integration (통합)을 사용하여 Fastmove는 application이 코드 변경 없이 DMA and 그것의 새로운 features (기능)을 투명하게 활용할 수 있도록 한다.
MySQL, GraphWalker and Filebench와 같은 standard benchmarks인 TPC-C와 PageRank와 같은 인기 있는 graph algorithm과 함께 NOVA, ext4-DAX 및 XFS-DAX 상에서 세 가지 data-intensive application을 실행한다. Single- and Multi-socket settings환경에서 conventional CPU-only NVM 접근과 비교하여 Fastmove는 TPC-C의 MySQL에서 peak throughput의 속도를 1.13 - 2.16 배 향상시키고, average latency를 17.7 - 60.8% 감소시키고 data movement에 사용되는 CPU 사용량을 37.1 - 68.9% 절약한다. 또한, GraphWalker를 통해 graph algorithm의 execution time을 39.7 - 53.4% 단축하고 Filebench의 throughput speed를 1.12 - 1.27 배 향상시킨다.