1 / 61

Evolution of the Atlas data and computing model for a Tier-2 in the EGI infrastructure

Evolution of the Atlas data and computing model for a Tier-2 in the EGI infrastructure. Álvaro Fernández Casaní on behalf of the IFIC Atlas computing group. II PCI2010 Workshop Valencia, 10 th -12 th January 2012. Outline. Introduction: IFIC and Atlas Tier2 SPAIN

imogene-pennington
Télécharger la présentation

Evolution of the Atlas data and computing model for a Tier-2 in the EGI infrastructure

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. Evolution of the Atlas data and computing model for a Tier-2 in the EGI infrastructure Álvaro Fernández Casaní on behalf of the IFIC Atlas computing group II PCI2010 Workshop Valencia, 10th-12th January 2012

  2. Outline • Introduction: IFIC and Atlas Tier2 SPAIN • Evolution of the Atlas computing model • Flattening the model to a Mesh • Tier 2 duties • Data distribution: Tier2s policy • Data shares ( site classification) • Availability and Connectivity • Networking • Dynamic data distribution and caching • Remote data access • Real Users example

  3. ATLAS CENTERS http://dashb-earth.cern.ch/dashboard/dashb-earth-atlas.kmz

  4. Introduction

  5. SPAIN CONTRIBUTION TO ATLAS

  6. Spanish ATLAS Tier2

  7. IFIC Computing Infrastructure Resources ATLAS • CSIC resources (not experiment resources): • 25% IFIC users • 25% CSIC • 25% European Grid • 25% Iberian Grid • To migrate Scientific application to the GRID EGI

  8. Last year summary • 3,579,606 Jobs • 6,038,754 CPU consumption hours • 13,776,655 KSi2K (CPU time normalised) • Supporting 22 Virtual Organizations (VOs)

  9. Prevision 2012For IFIC • For 2012: • We already fulfil the CPU requirements • Increasing Disk: • 230 TB => 4 SuperMicro x 57.6 (2 TB disks)

  10. EVOLUTION OF THE ATLAS COMPUTING MODEL

  11. Previous ATLAS cloud model • Hierarchical Model based on the Monarc network topology • CLOUDS OF Tier-1 with its geographical related Tier-2 • Possible communications: • T0-T1 • T1-T1 • Intra-cloud T1-T2 • Forbidden communications: • Inter-cloud T1-T2 • Inter-cloud T2-T2 Simone CampanaSoftware & Computing Workshop (April’11) https://indico.cern.ch/conferenceDisplay.py?confId=119169

  12. Shortcomings of Cloud Boundaries DATA FLOWS TOO STRICT OPERATIONAL PROBLEMS More info: Simone CampanaSoftware & Computing Workshop (April’11) https://indico.cern.ch/conferenceDisplay.py?confId=119169 • Consolidation of User Analysis outputs problematic • Analysis runs over many cloud • Consolidation needs to “hop” the data through the T1s • MonteCarlo production must confine one task to one cloud • To facilitate the output aggregation at T1 • Replication of datasets (PD2P) more inflexible • Need to replicate from T1 to T2s of the same cloud • Or “hop” through a T1 • T2s can not really be used as storage of “primary” data • Issues in creating secondary copies at other T1s • Tier-2 limited usability due to T1 downtimes • Dependency on LFC data catalog

  13. Solving issues LFC Spanish Cloud planned this week ! • Make Tier-2 activities more independent • Reduce service dependencies: • Move Catalog (LFC) to CERN: • Backup at US . Another at UK • Remove from tier-1 • Analysis jobs will be able to run during T1 downtime • Production jobs can keep running during T1 downtime • Benefit from technology improvements: • The network model today does not really resemble the Monarc model • Many T2s are connected very well with many T1s • Many T2s are not that well connected with their T1 • So it makes sense to break cloud boundaries.

  14. Flattening the model Network is a key component to optimize the use of the storage and CPU • to break cloud boundaries. • Let DDM freely transfer from every site to every site • Inter-cloud direct transfers • Multi-cloud production • Not quite there • Some links simply have limited bandwidth • In those cases, several hops will anyway be needed • Defining T2Ds is an attempt to break cloud boundaries “for the cases where it makes sense”

  15. Categorization of Sites: T2D • T2Ds are the Tier-2 sites that can directly transfer any size of files (especially large files) not only from/to the T1 of the cloud (to which they are associated to) but also from/to the other T1s of the other clouds. • T2Ds are candidates for • multi-cloud production sites • primary replica repository sites (T2PRR)

  16. ATLAS Tier2 Share Revision • In 2010, many Tier2s got full; T2 data distribution share revision • The data distribution to Tier-2 (and Tier-3g) sites should take into account • the network connectivity of the sites (thus, T2Ds should have more data) • the availability for analysis. • Introduced at Software & Computing Workshop 20 July 2011 • Up to ADC Ops to define the distribution policy among T2s: Preference to • Reliable sites for analysis (fraction of time with analysis queue online) • Well connected sites (T2Ds) to transfer datasets quickly • Started from summer 2011 • Based on T2D list defined on 1st July 2011 • Simplified with 4 groups treated equally : • Alpha (60% share-17 sites) : T2Ds with > 90 % reliability (60%) • Bravo (30 % share-21 sites) : non-T2Ds with > 90% reliability • Charlie (10 % share-12 sites) : Any T2 with 80%<reliability<90% • Delta (0% share-13 sites): Any T2 with reliability < 80 % • one single list both for pre-placement and dynamic data

  17. ATLAS Tier2

  18. Spanish Cloud Data shares Last Month http://atladcops.cern.ch:8000/drmon/crmon_TiersInfo.html IFIC IFIC is alpha T2D site -> also candidates for more datasets in PD2P ( see later)

  19. Availability • Hammercloud: for ATLAS • Distributed Analysis testing system • For avoiding jobs go to problematic sites • Can exclude sites if test jobs are not passed • EGIhas own different availability tools: • Based on ops VO • Can have conflicting results: • Last month IFIC was ATLAS ALPHA SITE ( >90 % availability) • But had just 64% ops egi availability due to a config issue LAST MONTH http://dashb-atlas-ssb.cern.ch/dashboard/request.py/siteviewhistorywithstatistics?columnid=564&view=shifter%20view

  20. Connectivity Transfer monitoring Inter-cloud direct transfers

  21. Services • User Interfaces: • UI00: (LVSDR) • UI04: SL 5 64 bits • UI05: SL 4 64 bits • UI06: SL 5 64 bits • Computing Elements (cream-ce) and Worker nodes : • WN (CE02): SL 5 64 bits Glite 3.2 with MPI and shared home in Lustre • WN (CE03): SL 5 64 bits Glite 3.2 (puppet) • WN (CE05): SL 5 64 bits Glite 3.2 (puppet)

  22. Job Distribution at Tier-2 PIC TIER-1 – 2012 More Analisys jobs at IFIC TIER 2 • Tier 2 usage in data processing • More job slots used at Tier2s than at Tier1 • Large part of MC production and analysis done at Tier2s • More ‘digi+reco’ jobs and “group production” at Tier-2s • less weight at Tier-1s • Production shares to be implemented to limit “group production” jobs at Tier-1, and run at Tier-2s • Analysis share reduced at Tier-1s

  23. Datadisk in tier2 February ‘12 Total Capacity (Usable/Margin) T2_datadisk Used (according SRM) T1_datadisk primary secondary Used (according dq2) 20PB http://bourricot.cern.ch/dq2/accounting/atlas_stats/0/ • Tier 2 usage disk space • T2_datadisk ≈ T1_datadisk in volume • T2_datadisk ≈ input for data processing (secondary replicas)

  24. For Spanish Sites February ‘12 http://bourricot.cern.ch/dq2/accounting/t2_spacetoken_view/SPAINSITES/datadisk/30/

  25. IFIC Storage resources • Based in SUN: • X4500 + X4540 • Lustre v1.8 • New Disk servers: • SuperMicro, SAS disks, capacity 2 TB per disk and 10 Gbe connectivity • Srmv2 (STORM) • 3 x gridftp servers

  26. Storage resources Pledges

  27. Storage resources

  28. IFIC Lustre Filesystems and Pools • Different T2 / T3 ATLAS pools , and separated from other Vos • Better Management and Performance • 4 Filesystems with various pools: • /lustre/ific.uv.es Read Only on WNs and UI. RW on GridFTP + SRM • /lustre/ific.uv.es/sw. Software: ReadWrite on WNs, UI (ATLAS USES CVMFS) • /lustre/ific.uv.es/grid/atlas/t3 Space for T3 users: ReadWrite on WNs and UI • xxx.ific.uv.es@tcp:/homefs on /rhome type lustre. Shared Home for users and mpi applications: ReadWrite on WNs and UI CVMFS FOR ATLAS • With the Lustre release 1.8.1, we have added pool capabilities to the installation. • Allows us to partition the HW inside a given Filesystem • Better data management • Assign determined OSTs to a application/group of users • Can separate heterogeneous disks in the future

  29. CERN CVMFS http://atlas-install.roma1.infn.it/atlas_install/usage_plots.php CVMFS: A caching, http based read-only filesystem optimised for delivering experiment software to (virtual) machines.

  30. CVMFS at IFIC (+SQUID) • Installed in all our WNs and UIs since September 2011 • Easy installation (only 2 configuration files) • 20 GB/repository • There is not a dedicated partition • Using the same SQUID as frontier (sq5.ific.uv.es) • the squid server is pointing to the public replicas (CERN, BNL, RAL) • The performance until now is so good. • We are monitoring SQUID via CACTI (SNMP) • Reduced job setup time • Better performance for analysis jobs • Reduced load on Lustre MDS Sept’2011 Sept’2011

  31. Storm SRM server Access like a local file system, so it can create and control all the data available in disk with a SRM interface. Coordinate data transfers, real data streams are transferred with a gridFTPserver in another physical machine. Enforce authorization policies defined by the site and the VO. Developed Authorization plugin to respect local file system with the corresponding user mappings and ACL's

  32. IFIC Network Recent Upgrade FTP servers to satisfy requirements for alpha sites • Cisco 4500 – core centre infrastructure. • Cisco 6500 – scientificcomputinginfrastructure • Data servers: • Sun with 1GB connection. • Channel bonding tests were made aggregation 2 channels • SuperMicro with10 Gbe • WNs and GridFTP servers with 1GB • 10 Gbe reach 1Gbit per data server Data Network based on gigabyte ethernet. 10GBe uplink to the backbone network

  33. Data Distribution: dynamic data distribution and caching

  34. Atlas presentation – Michael Ernst – BNL – LHCONE 12 May 2011, Washington

  35. PandaDynamicDataPlacement (PD2P) https://twiki.cern.ch/twiki/bin/viewauth/Atlas/PandaDynamicDataPlacement • resources utilization policy evolved during First year of data taking: • “jobs go to data” “data and jobs move to the available CPU resources”. • Also a dynamic data placement approach has been employed • Tier-1 algorithm • Primary copy at tier-1 basend on Planned data placement • Secondary copy when data popular, with location based on pledges • Tier-2 algoritm: • Jobs submitted to Panda trigger PD2P • Replicates popular datasets to a Tier2 with highest weight • More details: https://twiki.cern.ch/twiki/bin/viewauth/Atlas/PandaDynamicDataPlacement

  36. A further step: caching • PD2P makes data movement more dynamic and user-driven, but replication of dataset may suffer from other problems for the users: • Latency to finally access the data • Complexity of the tools • Still works on the idea of datasets replication • Explore the possibility to access to file or subfile dynamically without explicit replication • Xrootd allows federation of resources and to redirect the client to the data source. • Shorten latencies by caching in the xrootd server • To work correctly, there has to be efficient event data I/O with minimal transactions between application and storage

  37. PD2P in Es-Cloud last month • EsCloud T2Ds = IFIC and IFAE ES 166 IFIC 49 http://panda.cern.ch/server/pandamon/query?mode=pd2p&action=plots Es-Cloud Tier 2 sites getting more datasets Bit more that Tier-1 Pic

  38. User Example

  39. Example; Grid & Physics Analysis • Distributed Computing and Data Management Tools based on GRID technologies have been used by the IFIC physicists to obtain their results • As an example, the Boosted Top candidate presented by M. Villaplana

  40. Distributed Analysis in ATLAS • ATLAS has a specifically system for Production and Distributed Analysis (PANDA): • Including all ATLAS requirements • Highly automated • Low manpower • Unifies the different grid environments (EGI-Glite, OSG and EGI-ARC) • Monitoring web pages • Reference: http://panda.cern.ch

  41. Distributed Analysis in ATLAS • For ATLAS users, GRID tools have been developed: • For Data management • Don Quijote 2 (DQ2) • Data info: name, files, sites, number,… • Download and register files on GRID,.. • ATLAS Metadata Interface (AMI) • Data info: events number, availability • For simulation: generation parameter, … • Data Transfer Request (DaTri) • Users make request a set of data (datasets) to create replicas in other sites (under restrictions) • For Grid jobs • PanDa Client • Tools from PanDa team for sending jobs in a easy way for user • Ganga (Gaudi/Athena and Grid alliance) • A job management tool for local, batch system and the grid

  42. Tier2 and Tier3 examples from Spain • At IFIC the Tier3 resources are being split into two parts: • Resources coupled to IFIC Tier2 • Grid environment • Use by IFIC-ATLAS users • Resources are idle, used by the ATLAS community • A computer farm to perform interactive analysis (proof) • outside the grid framework • Reference: • ATL-SOFT-PROC-2011-018

  43. Daily user activity in Distributed Analysis • An example of Distributed Analysis in heavy exotic particles • Input files • Work flow:

  44. Daily user activity in Distributed Analysis • Just in two weeks, 6 users for this analysis sent: • 35728 jobs, • 64 sites, • 1032 jobs ran in T2-ES (2.89%), • Input: 815 datasets • Output: 1270 datasets • 1) A python script is created where requirements are defined • Application address, • Input, Output • A replica request to IFIC • Splitting • 2) Script executed with Ganga/Panda • Grid job is sent • 3) Job finished successfully, output files are copied in the IFIC Tier3 • Easy access for the user

  45. User experience • PD2P is transparent to users, but eventually they learn that it is in action: • “a job was finally sent to a destination that did not originally had the requested dataset” • “I check later PD2P has copied my original dataset” • “ I just realized because I used dq2 previously to check where the dataset was” • Another question is that user datasets are not replicated: • “We see a failure because it is not replicating our homemade D3PDs”

  46. Summary

  47. BACKUP SLIDES

  48. A TIER 2 in ATLAS Main activities

  49. Spanish Tier2 Number(more info: Pepe’s slides) • The ATLAS Spanish Tier2 (T2-ES) consists in a federation of 3 Spanish Institutions (see Jose’s talk): • IFAE-Barcelona (25%) • UAM- Madrid (25%) • IFIC-Valencia (50%, coordinator) • The T2-ES represents 5% of the ATLAS resources (between 60-70 T2s): • References: • J. Phys. Conf. Ser. 219 072046

  50. Summary

More Related