Caching for Non-independent Content: Improving Information Gathering in Constrained Networks

This paper explores the design of in-network caching policies for non-independent spot reporting content. The problem is motivated by autonomous sensor devices, such as soldier mounted cameras, that collect and store correlated data from the physical world. Physical states exhibit correlations in space and in time that need to be considered in caching decisions. Prior work in sensor networks that exploits correlations between content items focused more on efficient data collection, not caching, whereas work on caching mostly assumed that cached objects are independent. In contrast, when cached objects overlap in information content, such as pictures of a scene taken from different vantage points, the utility from storing similar items is sub-additive, motivating a fresh look at content-aware replacement policies. We describe a new caching policy, called diversity caching, that factors in both popularity and non-independence among objects into cache replacement algorithms. In a military environment, such a caching scheme supports a multi-modal information access model. Edge devices collect more data than the terrestrial network can handle, push high priority content directly through the network, and then use delay tolerant store and forward mechanisms to transmit the remaining content upon return to base. In information fusion systems, this provides immediate access to a wide variety of user queried content while remaining within the constraints of the terrestrial network and not relying on more expensive satellite networks. This paper evaluates diversity caching in the context of a disaster response application, where survivors and first responders survey damage and up-load observations into caches that are queried by others. Evaluation shows that diversity caching substantially improves a notion of coverage over previous replacement policies, such as LRU and Intentional Caching.