When the network crumbles: an empirical study of cloud network failures and their impact on services
Proceedings of the 4th annual Symposium on Cloud Computing
Hi-index | 0.00 |
Cloud infrastructure providers (CIPs) usually build a large number of geo-distributed data centers to cater to the recent proliferation of Cloud Computing. The CIPs commonly use multiple internet service providers (ISPs) to interconnect their geo-distributed data centers. Diurnal patterns and leftover bandwidth were effectively exploited by past studies to improve bandwidth utilization and minimize cost on inter-data center traffic. However, the growing number of inter-domain traffic was neglected. Inter-domain communications may become the potential bottleneck of inter-data center bulk transfers. Moreover, the rising inter-domain traffic increases the ISP operational cost, which will be not beneficial to reduce the CIPs' bandwidth cost or improve the quality of services over a long run. In this paper, we present a scheduling scheme that considers both bandwidth utilization and ISPs friendliness via a store-and-forward mechanism. The problem is modeled on a time-expanded graph. We compared our scheme with general bulk transfers mechanism under several different simulation settings. The results demonstrate that our strategy reduces the inter-domain traffic tremendously and achieves the ISP-friendliness significantly.