Séminaire ICI : Yao Li
ENSEA, salle 331
Modeling (Coded) Chunk-Based Content Delivery by the Coupon Collector's Problem
In chunk-based content distribution, files are fragmented at the source and the fragments (chunks) are distributed individually throughout the network. The way the chunks are circulated in the network affects the overall throughput of the network.
Instead of distributing the original file chunks, nodes can send out linear combinations of the chunks they hold, and this results in a throughput increase. But for a number of practical concerns (e.g., computational complexity and synchronization), chunks are grouped into (possibly overlapping) sets known as generations, and only chunks within the same generation are allowed to be linearly combined. A price to pay for combining (coding) only within generations is throughput reduction. We analyze the effects of generation size on the throughput by modeling the collection of (coded) chunked content as a "coupon collector's brotherhood" problem. Further, we extend the results from the collector's brotherhood problem and characterize the impact of introducing random overlaps between generations on the throughput. It is shown that proper choice of generation overlaps can improve throughput without raising computational complexity.