The Impact of Replacement Granularity on Video Caching

In this paper the idea that large objects, such as video files, should not be cached or replaced in their entirety, but rather be partitioned in chunks and replacement decisions be applied at the chunk level is examined. It is shown, that a higher byte hit ratio (BHR) can be achieved through partial replacement. The price paid for the improved BHR performance is that the replacement algorithm, e.g. LRU, takes a longer time to induce the steady state BHR. It is demonstrated that this problem could be addressed by a hybrid caching scheme that employs variable sized chunks; the use of small chunks leads to the maximization of BHR in periods of stable video popularity, while large chunks are used when extreme popularity changes occur to assist the fast convergence to the new steady state BHR.