Reimplementing the Cedar File System Using Logging and Group Commit
(Hagmann)

This paper describes modifications, with ideas borrowed from database
systems, to the Cedar file system (CFS). These borrowed ideas include
logging and group commit which help improve performance and robustness.
Design goals for the file system, include; a description of the failure
model with assumptions of what sector failure rate should be tolerated;
performance expectations of near hardware speed; safety from software errors
(both from the implementation and from the system); quick recovery time in
case of system crash (since disk sizes were expected to continue growing,
the existing filesystem recovery model was expected to be too time
consuming); and the use of off-the-shelf hardware. These design goals seem
consistent with that of most file systems today.

'The principle design concept is locality on the disk...', where information
about files are kept in the file name table, this in contrast to the
original CFS where information about files are kept in the file name table
as well as headers and labels. The file name table and the leader pages of
files (used for checking proper operation) are stored in an in-memory log
and two copies of all log records are periodically written for redundancy,
with every page written on two different sectors of the disk. To reduce
writes to the disk, group commit batches updates for log writes. The data
portion of a file did not go through the logging mechanism, because the file
system supported versioning and most files would only be written once. The
VAM a bitmap describing free disk pages were kept in memory and
reconstructed at startup.

Preserving locality on the disk, delayed writes and keeping the VAM in
memory help improve directory lookups and file create performance. The
statement calling for 'drastic measures' to handle large records in the
description of the logging algorithm is worrisome ('trust me.'). While the
paper does describe interesting concepts there is a lot of repetition in the
presentation, which made for a confusing read. The implementation gives the
impression of a hack to conform to changes in hardware.