Benefits of high speed interconnects to cluster file systems: a case study with Lustre

Abstract

Abstract

Resilience Design Patterns: A Structured Approach to Resilience at Extreme Scale (V.2.0)

“…The shared storage may be a dual-hosted hard drive, a networked storage using RAID, or a distributed replicated block device (DRBD) [154]. Such solutions have been extensively used in HPC environments for critical system services, such as the job and resource manager (e.g., SLURM [192] and Sun Grid Engine (SGE) [175]) and the parallel file system MDS (e.g., Parallel Virtual File System (PVFS) [146] and Lustre [194]). • Active/hot-standby redundancy using a commit protocol for state replication has been implemented for some HPC job and resource managers as part of high availability cluster solutions, such as HA-OSCAR [119] with its commit protocol for OpenPBS [16].…”

“…An implementation of HA-OSCAR supported high availability clustering for two job and resource managers, OpenPBS [16] and SGE [175]). Parallel file system MDSs, such as Lustre [194], support high availability clustering as well. • Active/standby redundancy also plays a role in resilience for parallel applications in HPC environments.…”

“…Active/Standby redundancy is typically used for critical hardware or software systems in HPC environments. For example, power supplies, voltage regulators, the parallel file system MDS in Lustre [194] and the SLURM [192] job and resource manager are often implemented in an active/standby fashion. Dual-modular redundancy for error detection and failure compensation and triple-modular redundancy for error detection and correction and failure compensation [78], dual-redundant parallel file system MDS solutions [92] and dual-redundant mission-critical HPC systems (e.g., weather forecast).…”

“…Examples: The Active/Standby structural pattern is typically used for critical hardware or software systems in HPC environments. For example, power supplies, voltage regulators, the parallel file system MDS in Lustre [194] and the SLURM [192] job and resource manager are often implemented in an active/standby fashion.…”

PetaShare: A Reliable, Efficient and Transparent Distributed Storage Management System

“…There have been many efforts in parallel and distributed data management systems to provide large I/O bandwidth [4,16,23]. However, metadata management is still a challenging problem in widely distributed large-scale storage systems.…”