dcache/srm @ osg abhishek rana frank würthwein ucsd
TRANSCRIPT
dCache/SRM @ OSG
Abhishek Rana
Frank Würthwein
UCSD
QuickTime™ and aTIFF (Uncompressed) decompressorare needed to see this picture.
OSG@UCSD 12/16/04
Frank Würthwein 2
SRM-dCache @ fnal
QuickTime™ and aTIFF (Uncompressed) decompressor
are needed to see this picture.
QuickTime™ and aTIFF (Uncompressed) decompressor
are needed to see this picture.
40GB/hourCMS write/hour
this week
CDF reads/daythis year
20TB/day
OSG@UCSD 12/16/04
Frank Würthwein 3
CDF Center @ FNAL
For comparison:Tier-1 requirement as Specified in CMS Computing model is800MB/sec
A day in 2003
Typical month in 2004
900MB/s
OSG@UCSD 12/16/04
Frank Würthwein 4
SRM-dCache scalability
• IO scalability is essentially infinite.– Depends primarily on # of pools and IO per pool.
• Scalability limits lie elsewhere:– Large rate of client requests
• Tight loops that try to open many files will bring system into its knees.
– Large rate of metadata accesses• Limit # of files per directory• Do not allow user access to virtual filesystem
Applications need to avoid “stupid” behavior!
OSG@UCSD 12/16/04
Frank Würthwein 5
The “geek gap”
• SRM-dCache is as complex as it is powerful.– Requires well packaged “cookie cutter”
configurations to minimize knowledge required from site admins.
– Requires some thought about hardware:• How much infrastructure hardware per cluster?• Choose filesystem that doesn’t fragment.• Choose linux kernel that isn’t buggy.• Some performance tuning (p2p, cost function, “quotas”)
OSG@UCSD 12/16/04
Frank Würthwein 6
Deployed flavors
• As cache for mass storage (HPSS,Enstore,OSM, …)
• Virtualizing space on RAID servers– replication only for load balancing
• Virtualizing compute node space– N copies of every file; “lazy” replication– virtual RAID in addition to load balancing.
OSG@UCSD 12/16/04
Frank Würthwein 7
FNAL plans
• Space reservation & quotas
• authorization module
• MIS
• Integration, deployment, operations
OSG@UCSD 12/16/04
Frank Würthwein 8
Reservations & Quotas
• Quota exists as a hard partitioning of pools and types of files that may be loaded into those pools.
• SRM v2 type reservations are being retrofitted into SRM v1.1 by 3/2005.
OSG@UCSD 12/16/04
Frank Würthwein 9
Security & ACLs
• Cron based synch of dcache & VOMS exists today.
• Authorization module with callout to GUMS by 3/2005
OSG@UCSD 12/16/04
Frank Würthwein 10
MIS
• Interface to GRIS & GIIS works
• Plan of Action being developed by 2/2005.– Internal accounting schema must allow
efficient queries.– Must allow VO specific queries– Send info to MonALISA
OSG@UCSD 12/16/04
Frank Würthwein 11
Integration, deployment, operations
• ~1/2005: Deploy resilient dCache at CMS sites & start operations challenge from Tier-1 to Tier-2. (UCSD will allow some VOs on part of its cluster)
• ~3/2005: Re-deploy new version incl. reservations, security, MIS enhancements.
OSG@UCSD 12/16/04
Frank Würthwein 12
Concerns by fkw
• Application’s use of SRM-dCache– dCache specific scalability issues– “Expectation management”
• Devil’s in the details– Personally, I would proceed well planned &
carefully integrate a few sites at a time rather than a large free for all.