ETRI-Knowledge Sharing Plaform

ENGLISH

성과물

논문 검색
구분 SCI
연도 ~ 키워드

상세정보

학술대회 Azalea-Unikernel: Unikernel into Multi-kernel Operating System for Manycore Systems
Cited 9 time in scopus Download 2 time Share share facebook twitter linkedin kakaostory
저자
전승협, 차승준, 람닉, 정연정, 김진미, 정성인
발행일
201810
출처
International Conference on Information and Communication Technology Convergence (ICTC) 2018, pp.1096-1099
DOI
https://dx.doi.org/10.1109/ICTC.2018.8539634
협약과제
18HS2500, 매니코어 기반 초고성능 스케일러블 OS 기초연구 (차세대OS기초연구센터), 정성인
초록
As applications such as a big data processing require more CPUs, manycore systems with a large number of CPUs have been developed to meet this requirement. However, without the in-depth consideration of parallelism, only increasing the number of cores cannot provide performance scalability. Furthermore, the current monolithic operating systems (e.g., Linux) cannot also provide performance scalability in manycore systems due to the cache coherency of shared data problems. As a result, multi-kernel operating systems have emerged as an alternative solution for scalability in manycore systems. We have developed a multi-kernel operating system called Azalea, which consists of a full-weight kernel (FWK) and lightweight kernel (LWK). The FWK handles heavy kernel services (i.e., file system services), and the LWK supports the minimal kernel functions as much as needed of application execution and eliminates the sharing of kernel data. However, there remain kernel noises in LWK such as context switching and page fault handling, even though they are less than Linux. In this paper, we propose Azalea-unikernel, which applies unikernel techniques into the LWK to reduce kernel noises. It eliminates privilege switching and address space switching by integrating user-kernel-address space. In particular case, the azalea-unikernel shows 7.5× better performance than LWK and Linux.
KSP 제안 키워드
Address space, Big Data Processing, Fault handling, File System, Kernel function, Many-core systems, Multi-kernel, Page fault, Performance and scalability, Reduce kernel, Shared data