People on Vidyo: Alessandra Forti, Andrea Sciaba, Andrew Melo, Bo Jayatilaka, Carlos Perez Dengra, Duncan Rand, Elizabeth Sexton-Kennedy, Edgar, emmanouil, Frederic Derue, Frank Wuerthwein, Ilija Vukotic, Julia Andreeva, Justas Balcas, Kaushik De, Laurent Duflot, Marcelo Vilaca, Markus Schulz, Michel Jouvin, Nikolai Hartmann, Oxana Smirnova, Riccardo Di Maria, Stephane Jezequel, Xavier Espinal
* News: no news.
* Alessandra Forti (University of Manchester (GB)) - Impact of Data Lake Model on Total Cost of Ownership : Manchester as ATLAS Large T2
- current setup details
- LHCb, DUNE, SKA currently share the storage
- UK is consolidating the storage
- smaller and medium sites will become CPU
- sites that lose storage because of manpower reduction tend to disappear and adding a cache is a second order problem
- assuming they don’t disappear they will put an extra load on the network and storage of the larger sites and the T1
- networking: 10Gbs redundant directly to the backbone, regularly saturated
- working on increasing bandwidth to backbone to 40Gbs
- changing storage solution from DPM
- JBOD and Erasure Coding seems the way to go (at the cost of more IOPS and bandwidth)
- future developments should keep separated the storage layer from the grid layer and QoS intelligence layer
- Data carousel based on data at T1 being transferred: trade off of networking vs disk (Manchester processed 243 TB of RAW data)
- Other QoS: Caches (network has been bottleneck till now), old storage (intelligent use of this is definitely needed), and/or Erasure Coding (resilience testing)
- Manchester will likely evolve towards: Ceph+xrootd (cum-HTTP-TPC)+EC and internal networking being upgraded in view of that (not now, and not in a rush)
- long term testbed and evolution like for DOME and TPC
* Frederic Derue (Centre National de la Recherche Scientifique (FR)) - Impact of Data Lake Model on Total Cost of Ownership : GRIF as ATLAS Medium/small T2
- existing DPM storage federation in France (e.g. ALPAMED)
- GRIF a distributed Tier2 in Ile de France
- info about HR and Funding Agencies in slides
- details about infrastructure, Network storage end-points per supported VO, and available resources in slides
- short/middle term evolution of GRIF presented wrt HR, support to VOs, CEs and SEs
- DataLake implementation in ATLAS: Rucio, FTS and other components already included
- recent evolution for http-tic protocols, macaroons etc
- with the probable phasing out of DPM, another storage tech should be chosen
- remote access, protection with XCache in front of WN or in network
- hardware problem / lost files
- GRIF not in DOMA yet because of several reasons
- diskless site model not interesting for GRIF
- going towards storage consolidation
- CEs is not clear if will be reduced
- SEs would reduce incrementally in accordance with endpoints
- DPM: disk are failing from time to time, loosing files is more rare
- DPM problems: great reduction in term of maintenance of the service; difficulties in exposing storage resources spread geographically (endpoints could not work in resilience mode)
There are minutes attached to this event.
Show them.
