Shared Data Spaces

In general, shared data space is a loose term for a system that facilitates shared state between distributed clients. Desirable characteristics of shared data spaces include scalability, reliability, and efficient propagation of updates. By this definition, some of the systems evaluated in this chapter are shared data spaces. The basic responsibilities of a shared data space are to maintain share state, handle client-side updates, and propagate updates to clients to provide data consistency.

In particular, we are concerned with a shared data space that supports Web-based DCAs. We propose an implementation of a shared data space that combines a number of the features of these other systems. Each feature contributes to the performance of a Web-based DCA through an aspect that relates to the Web environment.

• Fine-grained objects - The fundamental unit of data in a DCA, fine-grained objects propagate more efficiently than memory pages or disk blocks. Smaller, faster transmissions are good for keeping network traffic to a minimum.
• Network-wide name space - The ability to find an object anywhere on the Web requires a uniform network-wide name space.
• Caching - Users often access the same data repeatedly. Caching brings that data to the user's machine because accessing local data is faster and cheaper than accessing it remotely. Local state also enables the client to run disconnected or with unreliable connections. Caches also have the potential to keep client applications small in size. Small client applications with caches containing a minimal working set of data make it possible to run on computers with limited resources, e.g., PDAs or network computers.
• Unsynchronized access - Data access without the use of locks improves response time and simplifies implementation. In graphical DCAs, it also avoids the problem of denying a user access to an area of the screen just because someone else is holding a lock on an object in that area.
• Update-based Protocol - Collaboration implies that users want to see the contributions of others. Updates should be pushed into the user's client application as opposed to having the user or client application pull them in. Eagerly propagating updates will make data presentation more accurate.
• Global ordering of updates - This ordering of updates will be used to guarantee eventual consistency in a client/server system.
• Eventual consistency - Users can tolerate short-lived inconsistencies as long as they are guaranteed eventual consistency, i.e., all caches converge over time. Observing transient inconsistencies in data visualization will most likely appear as flickers on the screen due to the difference in computer and human response times.
• Client/server structure - A well-known location for connection to the server is essential. The server simplifies group management and encapsulates common system functionality, potentially minimizing the size of client applications. In addition, a hierarchy of servers makes the system scalable.
• Disconnected operation - Mobile computing requires disconnected operation and/or the ability to operate with unreliable connections. Also, many Internet Service Providers charge for connection time; disconnected operation saves money.

Encapsulating these techniques and properties into one data substrate will give us an efficient, scalable, and robust mechanism for sharing data between client applications over the Web.