Dynamic bin packing with heterogeneous dependent bins for Regionless in geo-distributed clouds

Cloud service providers use geo-distributed datacenters to provide resources and services to clients located in different regions. However, uneven population density leads to unbalanced development of geo-distributed datacenters and cloud service providers face a shortage of land resources to further develop datacenters in densely populated regions. Thus, it is a real challenge for cloud service providers to meet the increasing demand from clients in affluent regions with saturated resources and to better utilize underutilized data centers in other regions. To address this challenge, we study an online resource allocation problem in geo-distributed clouds, whose goal is to assign each user request upon arrival to an appropriate geographic cloud region to minimize the resulting peak utilization of resource pools with different cost coefficients. To this end, we formulate the problem as a dynamic bin packing problem with heterogeneous dependent bins where user requests correspond to items to be packed and heterogeneous cloud resources are bins. To solve this online problem with high uncertainty, we propose a simulation based memetic algorithm to generate robust offline proactive policies based on historical data, which enable fast decision making for online packing. Our experiments based on realistic data show that the proposed approach leads to a reduction in total costs of up to 15% compared to the current practice, while being much faster for decision making compared to a popular online method.