Transfer Learning for Cross-Model Regression in Performance Modeling for the Cloud

Authors: Francesco Iorio , Ali Hashemi , Michael Tao , Cristiana Amza

Venue: CloudCom 2019

publication url

Abstract

Performance characteristics of a complex system in different configurations are expensive to obtain due to the cost of sampling the system performance. We introduce ModelMap, a novel transfer learning technique for the performance modeling of configurable systems. ModelMap captures many explicit and latent types of dynamic system evolution, including configuration changes, scaling and hardware upgrades, by deriving and modeling directly these kinds of incremental transformations between system and/or application instances, over time. Modeling these transformations allows us to build accurate models for new configuration instances with just a few samples, and to interpolate across legacy models to build new models with no new samples at all. We experimentally test our method on a variety of system performance modeling and optimization scenarios. Compared to using conventional direct and incremental modeling techniques, our method achieves higher accuracy by up to an order of magnitude when the sampling budget is extremely limited, in particular between 0% to 5% of an exhaustive sampling budget. We also show how our method can be used to quickly derive an accurate resource allocation split that optimizes a given overall performance goal for co-hosted applications in a virtualized environment.