Proceedings of the 2015 ACM HPDC International Workshop on Scientific Cloud Computing (ScienceCloud), Portland, OR, June 2015.
Multi-tenant cloud infrastructures are increasingly used for high-performance and high-throughput domain science applications. Various cloud platforms, such as OpenStack and Amazon EC2, along with research efforts, such as NSF GENI and FutureGrid have attracted scientists to these platforms with the promise of virtually infinite compute resources.
This paper presents work toward providing better resource allocation accounting in multi-tenant cloud environments by understanding the subtle interference between network, compute, and storage resources. The experiments provide insight that help cloud administrators know how to best distribute virtual cores to physical cores considering the effect of advanced virtual network technologies on remote block I/O performance.
The results show that SR-IOV network interfaces to an SSD iSCSI device can provide extremely fast block I/O with minimal CPU overhead and minimal performance interference between tenants. In addition, careful mapping of virtual computation to physical computational cores is critical to increasing performance isolation.