SkyML: Infra-less machine learning on any cloud

SkyML: Infra-less machine learning on any cloud

In this talk we will introduce SkyML, a project being developed at UC Berkeley's Sky Computing Lab for "infra-less" machine learning (ML) on any cloud. SkyML offers ML practitioners and researchers ease of use, performance, and cost saving. For ease of use, SkyML abstracts away the cloud and infrastructure on which an ML application, such as training or hyperparameter tuning, is run. For performance, SkyML allows the easy provisioning and usage of best-of-breed accelerators across clouds. For cost saving, SkyML offers support for spot instances and automatic cleanup of resources when they become idle. We'll describe the current system and share our experience on its usage with early users.