Marrying sensors and innovative industries (for example, think aircraft engines instrumented for real-time preventive maintenance) leads to the creation of a huge time series of data samples. One appealing approach to efficiently managing this flood of data is to take advantage of cloud storage and computing platforms. Even though current infrastructures are rather large, the focus on efficient data handling remains paramount for both performance and cost matters.

The paper suggests a novel way to use lossy data compression to better manage these resources. More specifically, the idea is to cluster different but somewhat correlated data from all managed streams (for example, values from temperature gauges located close to one another or sequences of values from slow-varying sensors), create dedicated dictionaries for those chunks of data samples, and then use them to compress the stream clusters. The authors describe how the phases of data chunk generation and compression present in this new approach can be scaled to cloud-like proportion by leveraging the now-standard MapReduce paradigm. Experiments over terabyte-sized meteorological datasets show that significant cost, space, and time savings can be expected.

The main problem with this paper, as far as I can see, is that it has not been carefully prepared and edited: the approximate English makes the ideas and somewhat formal derivations difficult to understand, and the algorithms are at best glorified extracts of Java code, with too many inaccuracies. I cannot recommend reading this paper; skimming its abstract should be enough to get a feel for the ideas.