A Lightweight UID Mapping Scheme for the Lustre
Transcripción
A Lightweight UID Mapping Scheme for the Lustre
A Lightweight UID Mapping Scheme for the Lustre Filesystem Joshua Walgenbach So?ware Engineer, Data Capacitor Project Indiana University [email protected] Lustre • • • • Distributed Network Filesystem Separates Metadata from Data Data Scales Horizontally Works Great as a Filesystem for Clusters Lustre Scalable Object Storage MDS metadata server Client OSS object storage server Workflow is Not Local • High speed networks means you can collect data somewhere and analyze it across the country, or the world. • The tradiPonal method is copying data from one point to another. • Need centralized filesystem to manage workflow. Data Ownership • UNIX tracks file ownership with simple integers (UIDs) • Obviously won’t work across organizaPons • Need a method to map users into a single UID space IU UID Mapping Lightweight Not everyone needs / wants kerberos Not everyone needs / wants encrypPon Only change MDS code Want to maximize clients we can serve Simple enough to port forward UID Mapping • Lookup calls pluggable kernel module – Binary tree stored in memory – Based on NID or NID range – Remote UID mapped to EffecPve UID • Other lookup schemes are possible Announced One Year Ago • • • • • IU Lustre WAN Service Dedicated system of 360 TB for research Serve on a per project basis Single DDN S2A9950 Controller 6 Dell 2950 Servers – 4 Object Storage Servers – 1 pair of Metadata Servers for failover IU’s Lustre WAN on TeraGrid • ProducPon mount at PSC – AlPx 4700 and Login Nodes – GridFTP and Gatekeeper nodes • Test mounts at LSU, TACC, SDSC – Login nodes – LSU and SDSC ready to deploy on some compute nodes • Future plans for NCSA, ORNL, and Purdue EVIA – University of Michigan University of Michigan Uncompressed Digital Video IU HPSS SAMBA IU Windows Tools CReSIS – University of Kansas Greenland Ice Sheet Data From University of Kansas IU HPSS IU’s Quarry Supercomputer 3D Hydrodynamic Workflow HPSS PSC AlPx 4700 IU Astronomy Department Lustre Across the WAN IU Client Server PSC 17 ms latency Metadata Server Object Storage Servers Network IU ‐ PSC 2007 Bandwidth Challenge: Five ApplicaPons Simultaneously • AcquisiPon and VisualizaPon – Live Instrument Data • Chemistry – Rare Archival Material • HumaniPes • AcquisiPon, Analysis, and VisualizaPon – Trace Data • Computer Science – SimulaPon Data • Life Science • High Energy Physics Data acquisiPon from a global consorPum of crystallography faciliPes Post processed images of the palm leaves Sample images of the palm leaf of Sarvamoola granthas illustrating the performance of the image processing algorithms. (a) Stitched 8 bit grayscale image without normalization and contrast enhancement, (b) Final image after contrast enhancement SimulaPon of a High Energy ReacPon Performance Analysis with Vampir Alzheimer’s Amyloid PepPde Analysis Challenge Results Many Thanks • Stephen Simms, JusPn Miller, Nathan Heald, James McGookey,, Scom Michael, Mam Link (IU) • Kit Westneat (DDN) • Sun/CFS support and engineering • Michael Kluge, Guido Juckeland , Robert Henschel, Mamhias Mueller (ZIH, Dresden) • Thorbjorn Axellson (CReSIS) • Craig Prescom (UFL) • Ariel MarPnez (LONI) • Greg Pike and ORNL • Doug Balog, Josephine Palencia, and PSC Support for this work provided by the NaPonal Science FoundaPon is gratefully acknowledged and appreciated (CNS‐0521433)