PhEDEx: a novel approach to robust Grid data management Tim Barrass Dave Newbold and Lassi Tuura All Hands Meeting, Nottingham, UK 22 September 2005
Tim Barrass, Bristol, What is PhEDEx? A data distribution management system Used by the Compact Muon Solenoid (CMS) High Energy Physics (HEP) experiment at CERN, Geneva Blends traditional HEP data distribution practice with more recent technologies Grid and peer-to-peer filesharing Scalable infrastructure for managing dataset replication Automates low-level activity Allows manager to work with high level dataset concepts rather than low level file operations Technology agnostic Overlies Grid components Currently couples LCG, OSG, NorduGrid, standalone sites
Tim Barrass, Bristol, The HEP environment HEP collaborations are quite large Order of 1000 collaborators, globally distributed CMS is only one of four Large Hadron Collider (LHC) experiments being built at CERN Typically resources are globally distributed Resources organised in tiers of decreasing capacity Tier 0: the detector facility Tier 1: large regional centres Tier 2+: smaller sites-- Universities, groups, individuals… Raw data partitioned between sites, highly processed ready-for-analysis data available everywhere LHC computing demands are large Order 10 PetaBytes per year created for CMS alone Similar order simulated Also analysis and user data
Tim Barrass, Bristol, CMS distribution use cases Two principle use cases- push and pull of data Raw data is pushed onto the regional centres Simulated and analysis data is pulled to a subscribing site Actual transfers are 3rd party- handshake between active components important, not push or pull Maintain end-to-end multi-hop transfer state Can only clean online buffers at detector when data safe at Tier 1 Policy must be used to resolve these two use cases
Tim Barrass, Bristol, PhEDEx design Assume every operation is going to fail! Keep complex functionality in discrete agents Handover between agents minimal Agents are persistent, autonomous, stateless, distributed System state maintained using a modified blackboard architecture Layered abstractions make system robust Keep local information local where possible Enable site administrators to maintain local infrastructure Robust in face of most local changes Deletion and accidental loss require attention Draws inspiration from agent systems, autonomic and peer-to-peer computing
Tim Barrass, Bristol, Transfer workflow overview
Tim Barrass, Bristol, Production performance
Tim Barrass, Bristol, Service challenge performance
Tim Barrass, Bristol, Future directions Contractual file routing Cost-based offers for a given transfer Peer-to-peer data location Using Kademlia to partition replica location information Semi-autonomy Agents governed by many small tuning parameters Self modify- or use more intelligent protocols? Advanced policies for priority conflict resolution Need to ensure that raw data is always flowing Difficult real-time scheduling problem
Tim Barrass, Bristol, Summary PhEDEx enables dataset level replication for the CMS HEP experiment Currently manages 200TB+ of data, globally distributed Real life performance of 1 TB per day sustained per site Challenge performance of over 10TB per day Not CMS-- or indeed HEP-- specific Well-placed to meet future challenges Ramping up to get to O(10)PB per year TB per day Data starts flowing for real in the next two years
Tim Barrass, Bristol, Extra information PhEDEx and CMS : feel free to subscribe! CMS Computing model Agent frameworks JADE DiaMONDs FIPA Peer-to-peer Kademlia Kenosis Autonomic computing General agents and blackboards Where should complexity go? Agents and blackboards
Tim Barrass, Bristol, Issues Most issues fabric-related Most low level components experimental or not production-hardened Tools typically unreliable under load MSS access a serious handicap PhEDEx plays very fair, keeping within request limits and ordering requests by tape when possible Main problem is keeping in touch with the O(3) people at each site involved in deploying fabric, administration &c
Tim Barrass, Bristol, Deployment 8 regional centres, 16 smaller sites 110TB, replicated ~twice 1 TB per day sustained On standard Internet
Tim Barrass, Bristol, Testing and scalability
Tim Barrass, Bristol, PhEDEx architecture