AI

Dynamic iSCSI at Scale: Remote Paging at Google

Abstract

Google is experimenting with remotely mounting jobs’ binaries and data packages. Each package is served as an ext4 filesystem atop an iSCSI-mounted block device fed through dm-multipath and dm-verity. A typical job might run on O(1K) machines, employing a N-to-1 multipath configuration yielding O(10K) short-lived iSCSI sessions. Google thus sees O(100K) iSCSI sessions set up and torn down every minute, a rather atypical iSCSI deployment. We will explore the challenges presented by this deployment at scale.