TY - GEN
T1 - Failure-atomic synchronization of memory mapped data in non-volatile memory based system
AU - Haga, Koki
AU - Yamazaki, Kenichi
N1 - Funding Information:
This work was supported by NTT DOCOMO grant.
Publisher Copyright:
© 2018 IEEE.
Copyright:
Copyright 2019 Elsevier B.V., All rights reserved.
PY - 2018/3
Y1 - 2018/3
N2 - Thanks to the advancements in the research and development of high-speed nonvolatile memory (NVM) in recent years, the possibility has emerged that all hard disks and DRAM could be replaced by NVM. In this paper, we consider a computer with an architecture in which hard disks and DRAM are all replaced by NVM. In a conventional architecture, to perform a write operation to a file on an HDD data is copied to memory, modified, and written back to the storage. If the same process is followed in the NVM architecture, the high speed of NVM cannot be fully utilized. For this reason, in the NVM architecture files are mapped by mmap to virtual memory, and written directly as data on the NVM. On the NVM architecture, because writing to the mmaped area immediately takes effect on the file, there is no need to use msync. However, in such a case the consistency of the data cannot be guaranteed at a time of failure. In order to guarantee the consistency, we implemented a mechanism to create a copy of the file to be written. Writing is performed on the copy, and the copy becomes the original through a proposed atomic synchronization operation. In order to perform atomic synchronization, we implemented a mechanism to record journals. This paper describes the design and implementation in detail, and finally provides a performance evaluation.
AB - Thanks to the advancements in the research and development of high-speed nonvolatile memory (NVM) in recent years, the possibility has emerged that all hard disks and DRAM could be replaced by NVM. In this paper, we consider a computer with an architecture in which hard disks and DRAM are all replaced by NVM. In a conventional architecture, to perform a write operation to a file on an HDD data is copied to memory, modified, and written back to the storage. If the same process is followed in the NVM architecture, the high speed of NVM cannot be fully utilized. For this reason, in the NVM architecture files are mapped by mmap to virtual memory, and written directly as data on the NVM. On the NVM architecture, because writing to the mmaped area immediately takes effect on the file, there is no need to use msync. However, in such a case the consistency of the data cannot be guaranteed at a time of failure. In order to guarantee the consistency, we implemented a mechanism to create a copy of the file to be written. Writing is performed on the copy, and the copy becomes the original through a proposed atomic synchronization operation. In order to perform atomic synchronization, we implemented a mechanism to record journals. This paper describes the design and implementation in detail, and finally provides a performance evaluation.
KW - Fault tolerance
KW - File system
KW - High-speed nonvolatile memory
KW - Operating system
UR - http://www.scopus.com/inward/record.url?scp=85071596481&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85071596481&partnerID=8YFLogxK
U2 - 10.1109/SEATUC.2018.8788878
DO - 10.1109/SEATUC.2018.8788878
M3 - Conference contribution
AN - SCOPUS:85071596481
T3 - Proceedings - 12th SEATUC Symposium, SEATUC 2018
BT - Proceedings - 12th SEATUC Symposium, SEATUC 2018
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 12th South East Asian Technical University Consortium Sysmposium, SEATUC 2018
Y2 - 12 March 2018 through 13 March 2018
ER -