next up previous contents index
Next: File Sizes and Types Up: File Usage Previous: Operations on Directories   Contents   Index


File Sharing

Studies of file usage patterns in universities using a UNIX system [52, Ch.5.2.1] have shown that normally there is a low level of file sharing. The same studies have also shown that reads are much more common than writes. Putting these two observations together leads to the conclusion that there is a low level of write sharing! Furthermore, it seems that files often are read or written in their entirety, i.e., reads and writes are sequential rather than random.

At DIKU users have home directories, i.e., parts of the total directory structure are devoted to a single user. Normal use of the system is for the user to work alone on the files located in his or her home directory (and subdirectories); no sharing. If the files from a user's home directory are cached on a mobile computer used (only) by that user, then there will be no conflicts--it is unlikely that the user will be working on the same files both on the mobile computer and on the DIKU system at the same time. In Ficus [13] this behaviour is called a human write lock, and I think it is the main reason for the success of mobile computing systems using optimistic replica control strategies (see Section  3.3.3) such as Coda and Ficus.

So far, all studies of file usage patterns (at least to my knowledge) have been on UNIX like environments in settings similar to that at universities. If the same measurements where done in other environments then would they yield the same results? Databases for instance will undoubtedly have higher degrees of write sharing if the whole or big parts of (such as tablespaces) the database are looked upon as one file--file types and granularity of replication are discussed in Sections 3.1.5 and 3.2. Other environments with computer supported cooperative work (CSCW) may also show higher degrees of write sharing. What about users of mobile computers; do they automatically use their computers in the same way they would use stationary workstations?

Coda's assumptions are based on traces done on stationary workstations--how is one to know that users do not change their behaviour when working with their mobile computer? However, Coda has been successful [42].

In Bayou, on the other hand, it is a necessity that the applications are written with mobility in mind, i.e., that they are adapted to this new environment. Thus, it makes no assumptions with regards to the level of write sharing, but it does make assumptions about the ability of application developers to foresee possible conflicts--which I personally find reasonable!


next up previous contents index
Next: File Sizes and Types Up: File Usage Previous: Operations on Directories   Contents   Index

michael@garfield.dk
2000-10-13