Omid Pouya, BitFlood: Multicast Enabled P2P Data Sharing in Datacenters

Slides

Data transferring is considered to be a critical activity among different stages of a job in datacenters. Experiments on cluster computing applications such as MapReduce and Dryad indicates 33% to 50% of job completion times of these applications happens during the data transferring phases. In this project we target distribution or one-to-many communication patterns which occurs frequently in iterative optimization algorithms, code and VM distribution and fragment-replicate joins in Hadoop. Motivated by multicast capabilities of commodity switches in datacenters, we propose Bitflood, a Bittorent based protocol which (1) uses a multicast congestion control algorithm, (2) maintains tcp friendliness, and (3) improves the speed of data transferring while decreasing the network load.