1. d C a c h e Michael Ernst Patrick Fuhrmann Tigran Mkrtchyan d C a c h e M. Ernst, P. Fuhrmann, T. Mkrtchyan Chep 2003 Chep2003 UCSD, California

2. Overview dCache basics Modes of operation Attraction Model Dataset Location Management dCache Components (Overview) Native dCache access method (dCap) The GRID Storage Resource Manager (SRM) The Goal of the SRM Collaboration The Storage System Abstraction An SRM initiated transfer example M. Ernst, P. Fuhrmann, T. Mkrtchyan Chep 2003 Major SRM Topics

3. Used as an HSM Frontend, the dCache provides standard caching mechanisms to optimize tape accesses: Transfer speed adaption Tunable deferred HSM stores (space, time) Automatic staging Continuous garbage collection (no thresholds) Fetch ahead (from HSM) [in preparation] dCache : Modes of Operation M. Ernst, P. Fuhrmann, T. Mkrtchyan Chep 2003

4. Distributed Pool Architecture Topic Caches Tertiary Storage Enstore, OSM, … Super Cluster Caches (Experiment) Central Cache Host Caches Cluster Caches (Working Group) externally enforced attraction destination determined attraction

5. - dCache Pools without HSM backend can hold: Precious datasets Volatile datasets Files are never automatically removed. System can run out of disk space. Unused files are automatically removed. System won't run out of disk space. - The dCache can be operated in hybrid mode, running HSM and NON - HSM pools. dCache : Modes of Operation M. Ernst, P. Fuhrmann, T. Mkrtchyan Chep 2003

6. Typical Tier2 Today … GridFTP, bbcp Head/Storage Node Worker Nodes

7. Tier2 w/dCache dCache Admin Node GridFTP Worker Nodes

8. Tier2 w/dCache dCache GridFTP dCache GridFTP dCache/ Enstore Catalog GridFTP CASTOR FNAL CERN UCSD Florida, … Catalog

9. Client host IP or subnet numbers HSM groups Files are attracted (through store/retrieve requests) by Pools based on : Statically configured parameters, e.g. : Subdirectory Trees Dynamically taken parameters from system in operation, e.g. : Pool CPU cost (number of active movers) Pool Space costs (space left, age of datasets) M. Ernst, P. Fuhrmann, T. Mkrtchyan Chep 2003 dCache : The Attraction Model

10. dCache : Dataset Location Management With increasing load on pool nodes, the cache creates dataset replicas on less loaded nodes to smoothen hot spots, featuring Replica Redirection Services. Decreasing load marks dataset replicas for removal in case space is running short. Already staged files can be pinned to disk, overriding dynamic location management. Datasets can be defined 'sticky', independently of their status (CACHED, REPLICATED, ON TAPE ONLY) dCache : Dataset Location Management M. Ernst, P. Fuhrmann, T. Mkrtchyan Chep 2003

11. Pool Manager I/O Door Nodes SRM GFtp dCap (K)Ftp (Krb5,ssl) Http Admin File Name Space Database NFS Server pnfs File Name Space Provider OSM Enstore TSM ? Admin Doors dCache Components M. Ernst, P. Fuhrmann, T. Mkrtchyan Chep 2003 Pool Nodes HSMs

12. Besides supporting FTP, GridFTP, KerberosFTP and http, dCache has a native access protocol (dCap), allowing POSIX like file operations. dCache provides an implementation of the dCap protocol in C As shared object or preload library For Linux, Solaris and Irix Supporting automatic reconnect on network or server problems Win dCap.dll support Interfacing ROOT e.t.c. Providing security tunnels for Kerberos and ssl. dCache : Native Access Method (dCap) M. Ernst, P. Fuhrmann, T. Mkrtchyan Chep 2003

13. The Storage Resource Manager Collaboration In order to make local storage resources (e.g. disk space, tertiary storage space and large quantities of HEP information) globally available on Data Grids, a collaboration has been setup by LBNL, FermiLab, JLAB and CERN to define an abstract storage system interface, covering methods for : Storing and retrieving datasets Obtaining status information about datasets Pinning datasets (guarantee availability) Negotiating data transfer protocols. Defining dataset lifetimes and possible transitions. HEP sites intending to access remote/provide their storage resources on Data Grids can do so by integrating the SRM protocol into their local storage system. FermiLab is doing this for dCache. M. Ernst, P. Fuhrmann, T. Mkrtchyan Chep 2003

14. Storage System Abstraction OSMEnstoreTsm ? a s m i n e SRM HPSS dCache Cache HSM SRM Virtual Storage Layer J SRM SRM Client Store/Retrieve Pre-staging, Space allocation, Pinning Transfer Protocol Negotiation Data Transfer (e.g. GridFTP) Client System System Storage System Abstraction M. Ernst, P. Fuhrmann, T. Mkrtchyan Chep 2003

15. SRM Initiated Transfer Replica Catalog Any SRM map LFN -> SURL CERN- dataset/user/p/file1 srm://srm.fnal.gov/cern/user/p/file1 GET (SURL) srm://srm.fnal.gov/cern/user/p/file1 REPLY (SURL) REPLY (Request ID) #-12345 STATUS (Request ID) #-12345 REPLY (TURL) gridftp://pnfs/fnal.gov/cern/user/p/file1 GET (TURL) gridftp://pnfs/fnal.gov/cern/user/p/file1 FNAL SRM d C a c h e E n s t o r e STAGE

16. SRM Topics (SRM v2.1) Transfer Protocol Negotiation File Space Allocation / Reservation M. Ernst, P. Fuhrmann, T. Mkrtchyan Chep 2003 (includes direct I/O methods rfio,dCap, etc.) Dataset / space lifetime definitions (volatile, durable, permanent) SRM SRM third party transfers Directory support (mkdir / rmdir ) Security ( srm protocol will support gsi over http) Storing and retrieving datasets Obtaining dataset status information Pinning datasets (guarantee availability of data to job)

17. http://sdm.lbl.gov/srm-wg/ References dCache SRM www.dCache.org M. Ernst, P. Fuhrmann, T. Mkrtchyan Chep 2003