1. Storage Tank in Data Grid Shin, SangYong(syshin, #6468) IBM Grid Computing August 23, 2003

2. Storage Architecture Model - app. data is in files - files stored on block storage - all managed by storage mgmt s/w Application File System Block Virtualization Storage Devices Block subsystem Storage Management

3. SAN Block Virtualization Today SAN Block Virtualization  Common view of block storage  No server impact on storage change Emerging  No common view of block storage  Server impact on storage change - IBM block virtualization is Lodestone

4. High-end Host Application Midrange Disk array Shark, Brand X... FastT, Brand Y RAID Brick LVE = Lodestone Virtualization Engine LVE Disk array Functions Providing Virtual Disks Online- Dynamic Volume Sizing Advanced Copy Functions Economic Disaster Recovery Solutions. Different Level of Performance Data Backup with low price disk No Service Downtime etc LVE Extending Lodestone for Grid

5. File Systems - Current Capabilities Vs Grid Requirements SAN GPFS HPC, Engineering, Digital Media Access from servers in a cluster Concurrent multiple I/Os AIX and Linux OS only No access to other FS data Storage Tank Commercial, file sharing, DB serving Access from servers on SAN All servers and OSes No access to other FS data Grid requirements Access from any machine, any OS, anywhere Access to all file system data Planned Approach: Allow remote access to our file systems Provide multi-site support Integrate data from other sources

6. NFSv4 support for our file systems  We believe NFSV4 will be an important protocol for the grid has the necessary extensions for robust security and WAN access is the first NFS protocol to come through the standards process proposed standard in Dec. 2002; expected to be draft standard by 4Q03  Our plan is to provide NFSv4 support for our file systems (J2, GPFS and Storage Tank) Best case will be late 2004

7. Storage Tank (ST) - a SAN file system SN ST agent AIX ST agent Solaris Backup Data data Meta - Meta data ST agent Linux Capabilities: l access to ST data through Globus GridFTP interface l register ST files in Globus Replica Location Service l enabled to support OGSA services (e.g. replication)  centralized, policy-based storage management  cross-platform file sharing  performance comparable to local file system with direct client-to-storage data path ST agent Win2K ST Server GridFTP, NFS Prototypes: 2H02-1H03 Customer: CERN file attributes, file location info, control info LAN ST Clients

8. CERN Requirements  Data analysis of Large Hadron Collider (LHC) experiments Basic unit of data is an LHC event data represents physical collision between 2 protons 1 to few MBs stored within 1 GB files event metadata stored in an RDBMS  Tiered structure CERN is Tier 0 event data and metadata distributed to Tier 1 centers physicists at Tier 2 centers analyze data at Tier 1 centers  2.4 PB of disk and 14 PB of tape by 2007  Grid access (AFS/DFS like), simple storage management  IP SANs, not FC

9. Our Proposal  Use Storage Tank for basic storage infrastructure  Use iSCSI disks FAStT with iSCSI gateway or 200i  DB2 for event metadata  Research extensions NAS head for Storage Tank Grid access to Storage Tank Object Store prototype for disks

10. Single namespace across multiple sites - Replication of files for good performance - Extended protocols for consistency across replicas - Joint research w/ Johns Hopkins underway Tank NYC Fargo SFO Control Network (IP) SAN Meta-data Server Cluster Win2K ST Agent AIX ST Agent Solaris ST Agent Linux ST Agent data Integrated ST/NAS Appliance Meta-data Server Cluster Control Network (IP) SAN Meta-data Server Cluster Win2K ST Agent AIX ST Agent Solaris ST Agent Linux ST Agent data ST Extensions Prototype: 1H04 Customer: CERN, JHU Branch office Extend ST to Multiple Sites – Distributed Storage Tank

11. Ultimate Vision for Federated Grid File Systems Access Server Client... exporter Access Server Client... exporter Proxy Server Organization 1 Organization 2 file sources

12. Control Network (IP) SAN Meta-data Server Cluster Win2K ST Agent AIX ST Agent Solaris ST Agent Linux ST Agent data Control Network (IP) SAN Meta-data Server Cluster Win2K ST Agent AIX ST Agent Solaris ST Agent Linux ST Agent data Grid data repository GridFTP NFS NAS data repository Client NFS Extend ST to access data from other file systems/sources

13. Storage Management in Grid Computing Environment Applications Lodestone Storage TankShark, Tape, etc. CIM Provider Interface Storage Management Services CIM/XML OGSA OGSA-CIM Wrapper CIM/XML  IBM storage management products today (TSM, TSRM, ITSANM) and planned products (Merlot) cover a reasonable set of functions  We are converging, with the industry, on CIM/XML as the standard for storage device management  In support of grid, we expect: to convert our management solutions to Web/OGSA services to enhance functionality We are just starting to focus on grid implications for storage management

14. Block subsystem Summary of Data Grid Extend ST & GPFS Lodestone Support OGSA upper interface Support CIM lower interface OGSA CIM Application File System Block Virtualization Storage Devices Block subsystem Storage Management