Cümülön-D
Botong Huang, Jun Yang · Proceedings of the VLDB Endowment · 2017
We present a system called Cümülön-D for matrix-based data analysis in a spot market of a public cloud. Prices in such markets fluctuate over time: while users can acquire machines usually at a very low bid price, the cloud can terminate these machines as soon as the market price exceeds their bid price. The distinguishing features of Cümülön-D include its continuous, proactive adaptation to the changing market, and its ability to quantify and control the monetary risk involved in paying for a workflow execution. We solve the dynamic optimization problem in a principled manner with a Markov decision process, and account for practical details that are often ignored previously but nonetheless important to performance. We evaluate Cümülön-D's effectiveness and advantages over previous approaches with experiments on Amazon EC2.