SpinThrift: Saving energy in viral workloads

This paper looks at optimising the energy costs for data storage when the work load is highly skewed by a large number of accesses from a few popular articles, but whose popularity varies dynamically. A typical example of such a work load is news article access, where the most popular is highly accessed, but which article is most popular keeps changing. The properties of dynamically changing popular content are investigated using a trace drawn from a social news web site. It is shown that a) popular content have a much larger window of interest than non-popular articles. i.e. popular articles typically have a more sustained interest rather than a brief surge of interest. b) popular content are accessed by multiple unrelated users. In contrast, articles whose accesses spread only virally, i.e. from friend to friend, are shown to have a tendency not to be popular. Using this data, we improve upon Popular Data Concentration (PDC), a technique which is used to save energy by spinning down disks that do not contain popular data. PDC requires keeping the data ordered by their popularity, which involves significant amount of data migration, when the most popular articles keep changing. In contrast, our technique, SpinThrift, detects popular data by the proportion of non-viral accesses made, and results in lesser data migration, whilst using a similar amount of energy as PDC.