Abstract
This paper looks at optimising the energy costs for storing user-generated content when accesses are highly skewed towards a few "popular" items, but the popularity ranks vary dynamically. Using traces from a video-sharing website and a social news website, it is shown that the non-popular content, which constitute the majority by numbers, tend to have accesses which spread locally in the social network, in a viral fashion. Based on the proportion of viral accesses, popular data is separated onto a few disks on storage. The popular disks receive the majority of accesses, allowing other disks to be spun down when there are no requests, saving energy.
Our technique, SpinThrift, improves upon Popular Data Concentration (PDC), which, in contrast with our binary separation between popular and unpopular items, directs the majority of accesses to a few disks by arranging data according to popularity rank. Disregarding the energy required for data reorganisation, SpinThrift and PDC display similar energy savings. However, because of the dyamically changing popularity ranks, SpinThrift requires less than half the number of data reorderings compared to PDC.