sAccount: Resource Accounting of Shared Infrastructure in Multi-Tenant IT Platforms

In today’s consolidated IT platforms, the capability to accurately account the overall hardware resource usage among hosted applications can be necessary for a variety of resource management actions, and possibly also for auditing and explicit or implicit billing. We find that the increasing use of shared services in IT platforms, as exemplified by software-as-a-service offered by many clouds, renders existing solutions for accounting inadequate. We develop a resource accounting solution for a platform that consolidates multiple user applications and shared services on its server machines. We implement sAccount, a concrete realization of our solution, that consists of: (i) local monitoring facilities within the Xen hypervisor running on individual servers and (ii) a collective inference component that employs monitored data to derive accounting information for CPU and IO bandwidth resources. Using extensive evaluation on our prototype sAccount cluster involving a mix of synthetic and realistic (MySQL and HBase) shared services, we demonstrate the accuracy of sAccount. Our evaluation shows us that sAccount consistently offers less than 1% error in accounting at a low-overhead, whereas the error of a baseline technique employing state-of-the-art monitoring tools fluctuates between 5-150%. Furthermore, we also show that sAccount can enable administrators to improve the efficacy and agility of certain resource management tasks. For instance, sAccount allows correct and agile detection of the cause behind a CPU overload on servers running a shared MySQL database service. This, in turn, allows a simple resource control technique to correctly suppress the resource allocation for the source of this overload, whereas information offered by the baseline accounting leads to incorrect throttling decisions.