Underlying our work is the idea that in order for mobile computing to become the new standard model of computing, adaptive resource location and management will have to become an automatic function of distributed services software. The notion of constantly-networked, portable computers running modern operating systems is relatively new. Accordingly, we know of no work other than our own (already cited) on the topic of adaptive, dynamic mounting.
The Coda file system [Satyanarayanan90] supposes that mobile computing will take place in the form of ``disconnected operation,'' and describes in [Kistler91] a method in which the user specifies how to ``stash'' (read/write) files before disconnection and then, upon reconnection, have the file service run an algorithm to detect version skew. Coda can be taken as a point of contrast to our system, since the idea of disconnection is antithetical to our philosophy. We believe trends in wireless communication point to the ability to be connected any time, anywhere. Users may decide not to connect (e.g., for cost reasons) but will not be forced not to connect (e.g., because the network is unreliable or not omnipresent). We call this mode of operation elective connectivity.
An obvious alternative to our NFS-based effort is to employ a file system designed for wide-area and/or multi-domain operation. Such file systems have the advantages of a cache consistency protocol and a security model that recognizes the existence of many administrative domains. Large scale file systems include AFS [Howard88] and its spinoffs, Decorum [Kazar90] and IFS (Institutional File System) [Howe92]. Experiments involving AFS as a ``nation-wide'' file service have been going on for years [Spector89]. This effort has focused on stitching together distinct administrative domains so as to provide a single unified naming and protection space. However, some changes are needed to the present authentication model in order to support the possibility of a mobile client relocating in a new domain. In particular, if the relocated client will make use of local services, then there should be some means whereby one authentication agent (i.e., that in the new domain) would accept the word of another authentication agent (i.e., that in the client's home domain) regarding the identity of the client. Such could be made possible by cooperating Identification Servers [Johns93a,Johns93b]. For example, a client's RLP request could cause a server receiving it to call the client's identd server to find the identity of the user who initiated the request. The server may decide to deny the request if identification could not be made, or perhaps choose to ask another server (perhaps a master identification server back at the client's home base) for identity confirmation. Once an identity is confirmed, the server may log that information for future reference or for tracking in case of suspected break-in attempts.
The IFS project has also begun to investigate alterations to AFS in support of mobile computers [Honeyman91]. Specifically, they are investigating cache pre-loading techniques for disconnected operation and transport protocols that are savvy about the delays caused by ``cell handoff'' -- the time during which a mobile computer moves from one network to another.
Solaris 2.3's CacheFS [SMCC92b] allows for effective caching and synchronization of data between a client and NFS server. The main benefits of such a caching mechanism is the ability to use smaller, lighter, and less power-consuming disk drives -- especially important for mobile computers.
Plan 9's bind command has been designed to make it easy to mount new file systems. In particular, file systems can be mounted ``before'' or ``after'' file systems already mounted at the same point. The before/after concept replaces the notion of a search path. Plan 9 also supports the notion of a ``union mount'' [Pike91,Presotto92]. Several filesystem could be unified into one large one. Whenever files are identical, a client host might get parts of these files from any number of servers used to form the union. If one such server becomes inaccessible, but others still do, the client will continue to receive uninterrupted file service.
The Plan 9 bind mechanism is a more elegant alternative to our double mounting plus comparison. However, a binding mechanism -- even an unusually flexible one such as that of Plan 9 -- addresses only part of the problem of switching between file systems. The harder part of the problem is determining when to switch and whom to switch to.