Feeding the Masses: DNBD3. Simple, efficient, redundant block device for large scale HPC, Cloud and PC pool installations

Abstract

In computer center operations many sites operate large PC lecture pools or HPC clusters which can require similar or identical operating system images and software packages. Booting over the LAN allows instantaneously usable systems but requires the efficient provisioning of the root file system. Traditionally, general purpose file systems like NFS are used, but read-only Network Block Devices like the presented DNBD3 provide a range of attractive features, which can outperform alternatives across a range of situations. DNBD3 not only allows for caching and proxying at various levels, but it comes with a built-in performance monitor, versioning, and failover functionality. DNBD3 has been under development at Freiburg University for the past few years. It is released under a GPLv2 license, and consists of a Linux kernel module for the clients, and a user space executable for the servers. It is running in production for two highly heterogeneous use cases: as a distributed setup of campus-wide computer pools with more than 400 connected machines, and in the 1000+ node compute cluster backing the Freiburg HPC and Clouds. Aggressive local caching might even allow the use of mobile clients on WLAN infrastructures in stateless Linux operation.