Grid File Replication using SRM

1
Grid File Replication using Storage Resource
Management Presented By Alex Sim Contributors
JLAB Bryan Hess, Andy Kowalski Fermi Don
Petravick, Timur Perelmutov, Rich Wellner LBNL
Junmin Gu, Vijaya Natarayan, Ekow Otoo, Alex
Romosan, Alex Sim, Arie Shoshani WP2-EDG
Wolfgang Hoschek, Peter Kunszt, Heinz
Stockinger, Kurt Stockinger WP5-EDG
Jean-Philippe Baud http//sdm.lbl.gov/srm
2
How does file replication use SRMs(high level
view)

3
Main advantages of using SRMs for file
replication

• Can work in front of MSS
• To provide pre-staging
• To provide queued archiving
• Monitor transfer in/out of MSS
• Recover in case of transient failures
• Reorders pre-staging requests to minimize tape
  mounts
• Monitors GridFTP transfers
• Re-issues requests in case of failure
• Can control number of concurrent GridFTP
  transfers to optimize network use (future)
• SRMs role in the data replication
• Storage resource coordination scheduling
• SRMs do not perform file transfer
• SRMs do invoke file transfer service if needed
  (GridFTP)

4
Brief history of SRM since GGF4

• Agreed on single API for multiple storage systems
• Jlab has an SRM implementation based on SRM v1.1
  spec on top of JASMine
• Fermi Lab has an SRM implementation based on SRM
  v1.1 spec on top of Enstore
• WP5-EDG is proceeding with SRM implementation on
  top of Castor
• LBNL Deployed HRM-HPSS (which accesses files
  in/out of HPSS) at BNL, ORNL, and NERSC (PDSF)
• HRM-NCAR that accesses MSS at NCAR is in progress

5
SRM version 2.0

• Joint design and specification of SRM v2.0
• LBNL organized meeting to coordinate design, and
  summarized design conclusions
• SRM v2.0 spec draft version exists
• SRM v2.0 finalization to be done in Dec.
• Design uses OGSA service concept
• Define interface behavior
• Select protocol binding (WSDL/SOAP)
• Permit multiple implementations
• Disk Resource Managers (DRMs)
• On top of multiple MSSs (HRMs)

6
Brief SummarySRM main methods

• srmGet, srmPut, srmCopy
• Multiple files
• srmGet from remote location to disk/tape
• srmPut from client to SRM disk/tape
• srmCopy from remote location to SRM disk/tape
• srmRelease
• Pinning automatic
• If not provided, apply pinning lifetime
• srmStatus
• Per file, per request
• Time estimate
• srmAbort

7
Main Design Points

• Interfaces to all types of SRMs to be uniform
• Any Clients, Middleware modules, other SRMs
• Will communicate with SRMs
• Support a multiple files request
• set of files, not ordered, no bundles
• Implies queuing, status, time estimates, abort
• SRMs support asynchronous requests
• Non-blocking, unlike FTP and other services
• Support long delays, multi-file requests
• Support call-backs
• Plan to use event notification service
• Automatic Garbage Collection
• In file replication, all files are volatile
• As soon as they are moved to target, SRMs
  perform garbage collection automatically.

8
Current efforts on SRMs at LBNL

• Deployed HRM at BNL, LBNL (NERSC/PDSF), ORNL for
  HPSS access
• Developing HRM at NCAR for MSS access
• Deployed GridFTP-HRM-HPSS connection daemon
• Supports multiple transport protocols
• gsiftp, ftp, http, bbftp and hrm
• Deployed web-based File Monitoring Tool for HRM
• especially useful for large file replica requests
• Deployed HRM client command programs
• User convenience
• Currently developing web-services gateway
• Currently developing GSI-enabled requests

9
Web-Based File Monitoring Tool
10
HRMs in PPDG(high level view)
Replica Coordinator
HRM-COPY
HRM-GET
HRM (performs writes)
HRM (performs reads)
GridFTP GET (pull mode)
LBNL
BNL
11
HRMs in ESG(high level view)

12
File replication from NCAR to ORNL/LBNL
HPSScontrolled at NCAR
13
File Replication from ORNL HPSS to NERSC
HPSScontrolled at NCAR Non-Blocking Calls
14
Recent Measurements of large multi-file
replication
Event
time
Event
time
15
Recent Measurements of large multi-file
replication (GridFTP transfer time)
16
Recent Measurements of large multi-file
replication (Archiving time)
17
Recent Measurements of large multi-file
replication
Event
time
Event
time
18
Recent Measurements of large multi-file
replication (GridFTP transfer time)
19
Recent Measurements of large multi-file
replication (Archiving time)
20
File Replication from ORNL HPSS to NERSC
HPSScontrolled at NCAR
0 Request_Arrive_at_LBNL 1 Request_Arrived_at_ORNL
2 Staging_requested_at_ORNL 3 Staging_started_ORN
L 4 Staging_finished_ORNL 5 Callback_from_ORNL_to_
LBNL 6 GridFTP_Start_by_LBNL 7 Transfered_from_ORN
L_to_LBNL 8 Migration_Requested 9
Migration_Started 10 Migration_Finished 11
Notified_Client
HRM Client on NCAR HRM on LBNL and ORNL Max
Concurrent PFTPs5 Cache Size at LBNL30G Cache
Size at ORNL25G Max Concurrent Pinned File5 Max
Concurrent GridFTP3 GridFTP p4 GridFTP
bs1000000