Sonnet: A control-theoretic approach for resource allocation in cluster management

Cluster users expect to minimize the resource costs while ensuring target performance for different applications. It is particularly difficult to reach such a goal, because the applications are diverse with dynamic load changes, and interference exists between them. In addition, the performance of the applications depends on heterogeneous resources with different costs. However, existing works either use simplistic and generalized heuristics that disregard resource-specific characteristics or need suspending service to get expert knowledge to optimize the resource cost for a brand-new application or runtime, which fails to optimize the resource allocation finely. In this paper, we propose Sonnet, a control-theoretic approach to perform efficient resource allocation. Sonnet can efficiently optimize the cost of resources while satisfying the SLO by quickly establishing new application performance models through only online profiling and without affecting service. Experiments on Docker Swarm using various open-source benchmarks demonstrate that Sonnet can decrease the SLO violation rate by 91% while reducing resource costs up to 47% compared with the state-of-the-arts.