June 16-20, 2013

Leipzig, Germany

Presentation Details

Name: Storage & Data
(10) An Analysis of Lustre Metadata Server Scalability
Time: Monday, June 17, 2013
2:10 PM - 2:15 PM
Room:   Multi-Purpose Area 4 (MPA 4)
CCL - Congress Center Leipzig
Speakers:   Konstantinos Chasapis, University of Hamburg
Abstract:   The main form of storing data in current supercomputers is using files. Parallel file systems implement a file API to offer persistent storage to applications and use metadata for this purpose. Applications use files to store data in many forms, some of them store data in few huge files and others in many small files. Using many small files to store data, metadata file operations can dramatically affect the application performance since it will perform a vast amount of metadata operations.
The dominant file system in the top 10 list of supercomputers is Lustre. It is designed for extreme scalability at the I/O nodes but it offers only a single metadata server (MDS) resulting in a single point of conjunction for metadata operations. Recently a feature was added to support an additional passive metadata server for failover purposes to increase Lustre availability. Many metadata server designs have been proposed to improve Lustre metadata server architecture including clustered metadata, but as of version 2.3 of Lustre none of them have been implemented. The total amount of metadata occupies only a small percentage of the total address space so new storage technologies such as Solid State Drives (SSDs) that are more expressive than hard drives (HDDs) can be used as metadata server storage devices. SSDs are perfectly suited for metadata storage devices since they offer orders of magnitude more I/O operations per second (IOPS) that are mainly required to perform metadata operations.
In this poster we present an evaluation of Lustre metadata server scalability. The main aspects that we analyze are: scalability with the number of processes, the number of client nodes and the files per directory. Moreover we present a comparison between SSDs and HDDs to observe the actual benefit of SSDs. Finally, we also compare Lustre with Orion-Lustre that supports ZFS as a backend file system instead of the traditional ldiskfs.

Poster authors:
Konstantinos Chasapis, University of Hamburg; Michael Kuhn, University of Hamburg; Thomas Ludwig, DKRZ
Program may be subject to changes.