Ganger

Proceedings of the FREENIX track: 1999 USENIX Annual Technical Conference, June 1999, pp. 1-17

Marshall Kirk McKusick, Author and Consultant 
Gregory R. Ganger, Carnegie Mellon University 

Traditionally, filesystem consistency has been maintained across system failures either by using synchronous writes to sequence dependent metadata updates or by using write-ahead logging to atomically group them. Soft updates, an alternative to these approaches, is an implementation mechanism that tracks and enforces metadata update dependencies to ensure that the disk image is always kept consistent. The use of soft updates obviates the need for a separate log or for most synchronous writes. Indeed, the ability of soft updates to aggregate many operations previously done individually and synchronously reduces the number of disk writes by 40 to 70% for file- intensive environments (e.g., program development, mail servers, etc.). In addition to performance enhancement, soft updates can also maintain better disk consistency. By ensuring that the only inconsistencies are unclaimed blocks or inodes, soft updates can eliminate the need to run a filesystem check program after every system crash. Instead, the system is brought up immediately. When it is convenient, a background task can be run on the active filesystem to reclaim any lost blocks and inodes. This paper describes an implementation of soft updates and its incorporation into the 4.4BSD fast filesystem. It details the changes that were needed, both to the original research prototype and to the BSD system, to create a production-quality system. It also discusses the experiences, difficulties, and lessons learned in moving soft updates from research to reality; as is often the case, non-focal operations (e.g., fsck and ‘‘fsync’’) required rethinking and additional code. Experiences with the resulting system validate the earlier research: soft updates integrates well with existing filesystems and enforces metadata dependencies with performance that is within a few percent of optimal.

Download: pdf