Globus Data Services for Science Raj Kettimuthu Argonne National Laboratory/Univ. of Chicago Ann Chervenak, Rob Schuler USC Information Sciences Institute Globus Services for Data Intensive Science Data Movement: GridFTP (RFT) and Reliable File Transfer Service Replica management: Replica Location Service (RLS) and Data Replication Service (DRS)
New: Policy-based data placement service Access to databases and other data sources: OGSA Data Access and Integration (DAI) Service Talk Outline Examples of production data intensive science projects that use Globus services New features: GridFTP and RFT Replica management tools
Data placement services Data access and integration services The LIGO Project Laser Interferometer Gravitational Wave Observatory LIGO instruments in Washington State and Louisiana During science runs, produce up to 2 terabytes per day Published along with metadata at Caltech (archival site) Replicated at up to 10 other LIGO sites
LIGO scientists typically move data sets near to computational clusters at their sties The LIGO Data Grid Globus Services in the LIGO Data Grid Lightweight Data Replicator (LDR): data management system developed by LIGO researchers Globus data services: GridFTP: used for moving data around the Grid efficiently and securely Replica Location Service: catalogs deployed at all LIGO sites, keep track of locations of over 150 million files
Data Replication Service was developed to generalize the functionality in the LDR Other Globus services: Globus security Starting to deploy the Globus Monitoring and Discovery Service Earth System Grid objectives To support the infrastructural needs of the national and international climate community, ESG is providing crucial technology to securely access, monitor, catalog, transport, and distribute data in todays grid computing environment. HPC hardware running climate models ESG Portal ESG Sites 6 Bernholdt_ESG_SC07 ESG facts and figures Main ESG Portal 146 TB of data at four locations
1,059 datasets 958,072 files Includes the past 6 years of joint DOE/NSF climate modeling experiments IPCC AR4 ESG Portal 35 TB of data at one location 77,400 files Generated by a modeling campaign coordinated by the Intergovernmental Panel on Climate Change Model data from 13 countries 4,910 registered users 1,245 registered analysis projects Downloads to date Downloads to date 30 TB
106,572 files Worldwide ESG user base 245 TB 914,400 files 500 GB/day (average) IPCC Daily Downloads (through 7/2/07) > 300 scientific papers published to date based on analysis of IPCC AR4 data Slide Courtesy of Dave Bernholdt, ORNL ESG architecture and underlying technologies Metadata catalog
NcML (metadata schema) First Generation ESG Architecture Climate data tools OPenDAP-G (aggregation, subsetting) NCAR Cache NCAR MSS RLS SRM Data Mover Lite
Storage Resource Manager SRM Globus Security Infrastructure GridFTP Monitoring and Discovery Services Replica Location Service Access control
MyProxy User registration SRM DISK Cache SRM OPeNDAP-G SRM ESG Web Portal NERSC RLS Globus toolkit Security RLS MyProxy Data management
ORNL HPSS User Registration Catalogs Browsing Data Search Access Control Climate Metadata Data Download Data Subsetting Data Publishing Usage
Metrics RLS Monitoring Services Web Web Browser Browser Data Provider publish search browse download Web Web Browser Browser DML DML MSS, HPSS: Tertiary data storage systems
Data User LANL Cache GridFTP Data Transfers for the Advanced Photon Source One Australian user left nearly 1TB of data on our systems that we had been struggling to transfer via standard FTP for several weeks. The typical data rate using standard FTP was ~200 KB/s. Using GridFTP we are now moving data at 6 MB/squite a significant boost in performance! Brian Tieman Advanced Photon Source 30x speedup 9688 miles Whats New in Globus GridFTP and RFT Raj Kettimuthu Argonne National Laboratory and The University of Chicago
What is GridFTP? High-performance, reliable data transfer protocol optimized for high-bandwidth wide-area networks Based on FTP protocol - defines extensions for high-performance operation and security We supply a reference implementation: Server Client tools (globus-url-copy) Development Libraries
Multiple independent implementations can interoperate Fermi Lab and U. Virginia have home grown servers that work with ours. GridFTP Two channel protocol like FTP Control Channel Communication link (TCP) over which commands and responses flow Low bandwidth; encrypted and integrity protected by default Data Channel
Communication link(s) over which the actual data of interest flows High Bandwidth; authenticated by default; encryption and integrity protection optional Why GridFTP? Performance Cluster-to-cluster data movement Parallel TCP streams, optimal TCP buffer
Non TCP protocol such as UDT Order of magnitude greater Another order of magnitude Support for reliable and restartable transfers Multiple security options Anonymous, password, SSH, GSI Cluster-to-Cluster transfers Performance Mem. transfer between Urbana, IL and San Diego, 30 CA 25 20 15 10 Throughput (Gbit/s) 5 # Stream = 1 # Stream = 8 # Stream = 2 # Stream = 16
# Stream = 4 # Stream = 32 0 0 10 20 30 40 Degree of Striping 50 60 70 Performance 20 Disk transfer between Urbana, IL and San Diego, CA
15 10 5 Throughput (Gbit/s) # Stream = 1 # Stream = 8 # Stream = 2 # Stream = 16 # Stream = 4 # Stream = 32 0 0 10 20 30 40 Degree of Striping
50 60 70 Users HEP community is basing its entire tiered data movement infrastructure for the LHC computing Grid on GridFTP Southern California Earthquake Center (SCEC), European Space Agency, Disaster Recovery Center in Japan move large volumes of data using GridFTP An average of more than 2 million data transfers happen with GridFTP every day LOSF and Pipelining Traditional File Request 1 DATA 1 Pipelining File Request 1
File Request 2 File Request 3 DATA 1 ACK 1 ACK 1 File Request 2 DATA 2 DATA 2 ACK 2 File Request 3 ACK 2 DATA 3 ACK 3 DATA 3 ACK 3 Significant performance improvement for LOSF GridFTP over UDT
GridFTP uses XIO for network I/O operations XIO presents a POSIX-like interface to many different protocol implementations Default GridFTP GridFTP over UDT GSI GSI TCP UDT GridFTP over UDT Argonne to NZ Throughput in Mbit/s Argonne to LA Throughput in Mbit/s Iperf 1 stream
19.7 74.5 Iperf 8 streams 40.3 117.0 GridFTP mem TCP 1 stream 16.4 63.8 GridFTP mem TCP 8 streams 40.2 112.6 GridFTP disk TCP 1 stream 16.3 59.6 GridFTP disk TCP 8 streams
37.4 102.4 GridFTP mem UDT 179.3 396.6 GridFTP disk UDT 178.6 428.3 UDT mem 201.6 432.5 UDT disk 162.5 230.0 SSH Security for GridFTP
Port 22 Client sshd ROOT ssh Stdin/out GridFTP Server USER Multicast / Overlay Routing Enable GridFTP to transfer single data set to many locations or act as an intermediate routing node GridFTP with Lotman Client GridFTP Server
Lotman Reliable File Transfer Service (RFT) GridFTP client WSRF complaint fault-tolerant service RFT Client SOAP Messages Notificati ons (Optional) RFT Service Persistent Store CC CC GridFTP Server DC GridFTP Server
RFT - Connection Caching Control channel connections (and thus the data channels associated with it) are cached to reuse later (by the same user) RFT Service CC GridFTP Server CC DC GridFTP Server RFT - Connection Caching Reusing connections eliminate authentication overhead on the control
and data channels Measured performance improvement for jobs submitted using Condor-G For 500 jobs - each job requiring file stageIn, stageOut and cleanup (RFT tasks) 30% improvement in overall performance No timeout due to overwhelming connection requests to GridFTP servers Whats new in Data Access and Integration? Raj Kettimuthu on behalf of OGSA-DAI team What is OGSA-DAI? Middleware that allows data resources, such as relational or XML databases, to be accessed via web services What is OGSA DAI? OGSA-DAI executes workflows
OGSA-DAI is not just for data access, also does data updates, transformations and delivery. OGSA DAI Workflow Remote resource access OGSA-DAI data resource interaction Via a data resource plug-in Remote resource access Access a data resource managed by another OGSA-DAI server Remote resource access
Remote resource plug-in Basically a client to a remote OGSADAI server Runs queries via workflow submission Configured with URL of remote server Transparent to OGSA-DAI infrastructure Just another data resource plug-in OGSA-DAI 3.0 data sources OGSA-DAI data sources Resource for asynchronous data delivery
Data source service Web service Invoke GetFully via SOAP/HTTP Use WS-Addressing to specify data source ID data from workflow Expose Expose via via data data source source DataSource DataSourceService getFully() Client OGSA-DAI servlet
Data source servlet Invoke HTTP GET Use URL query string to specify data source ID data from workflow Expose Expose via via data data source source DataSource DataSourceRetrievalServlet HTTP GET Client
OGSA-DAI servlet Useful for service orchestration and job submission Taverna service-oriented workflow executor Taverna could submit workflow to OGSA-DAI OGSA-DAI returns URL Taverna passes URL as part of job to job submission service e.g. GRAM or GridSAM Data is pulled from the URL when the job is executed Advantages
Data is only moved when needed i.e. when the job executes Job execution components need no OGSA-DAI-specific components A join activity Virtual Organisations for Trials and Epidemiological Studies (VOTES) UK Medical Research Council project Relational databases Uses OGSA-DAI OGSA-DAI team developed join activities A join activity
SELECT id, x FROM tableOne ORDER by id SELECT myID, y FROM tableTwo ORDER by myID Run SQL SQL Run query query Run SQL SQL Run query query joinColumn1: id joinColumn2: myID Tuple merge join This is equivalent to running: SELECT id, x, y FROM tableOne, tableTwo where table1.id = table2.myID; Where tableOne and tableTwo are in two different databases
SQL views Imagine we have Patient and Doctor tables ID Name Age Sex ZIP 1 Ken 42 M IL1478305 456 2 Josie
F BN1 7QP 789 ID Name DN 123 Greene US-Chicago-G 456 Ross US-Chicago-R 789
Fairhead UK-Holby-F SQL CREATE VIEW command Define a DrPatient view to be 25 Dr ID SELECT p.id, p.name, p.age, p.sex FROM Patient p, Doctor d WHERE p.DrID = d.ID; Client runs SELECT * FROM DrPatient; Shorthand for complex queries Data access control e.g. staff with only access to the DrPatient view will be unable to access a patients ZIP OGSA DAI SQL views
Layer above the database to implement views Define views for databases to which you dont have write access Parses query Maps view to SQL query over actual database e.g if DrPatient was defined as SELECT p.id, p.name, p.age, p.sex FROM Patient p, Doctor d WHERE p.DrID = d.ID AND d.dn = $DN$; Can replace $DN$ by clients DN from their certificate provided using GT4 security components Doctors can only view their own patients Factor in the clients security credentials OGSA-DQP
Distributed query processing How it works Multiple tables on multiple databases are exposed to clients multiple tables in one virtual database Client is unaware of the multiple databases Databases can be exposed within one OGSA-DAI server or exposed by remote OGSA-DAI servers Query is parsed Query plan is created Query plan is executed each database has subqueries executed on it Results are combined Good for joins and unions
Whats new in data replication and placement services? Rob Schuler Objectives for Data Replication Improve Availability Safeguard against data inaccessibility due to network partition Improve Performance Safeguard against performance bottlenecks due to resource overload A Improve Durability Safeguard against data loss due to disk failure A A The Globus Replica Location Service
Distributed registry Records the locations of data copies Allows replica discovery RLS maintains mappings between logical identifiers and target names Must perform and scale well: support hundreds of millions of objects hundreds of clients Mature and stable component of the Globus Toolkit
Replica Location Indexes RLI RLI RLI LRC RLI LRC RLI LRC Local Replica Catalogs LRC New Features in RLS Embedded SQLite database for easier RLS deployment Open source relational database backends (MySQL, PostgreSQL) depend on ODBC libraries
Compatibility problems that have made DB deployment difficult Embedded DB back end now allows easy installation of RLS Allows easier evaluation of RLS by potential users SQLite offers good performance and scalability on queries Does not support multiple simultaneous writers, so not suitable for some high performance environments New Features in RLS Pure Java client implementation
Long-awaited Overcomes problems with JNI-based client, particularly on 64-bit platforms Improves reliability of portals that use RLS Java client Being used by several large applications (ESG, SCEC) WS-RLS interface: provides a WS-RF compatible web services interface to RLS Easier integration of RLS services into GT4 Web service environments Data Placement Services: Motivation
Scientific applications often perform complex computational analyses that consume and produce large data sets The placement of data onto storage systems can have a significant impact on Computational and storage resources distributed in the wide area performance of applications reliability and availability of data sets We want to identify data placement policies that distribute data sets so that they can be staged into or out of computations efficiently replicated to improve performance and reliability Data Placement and
Workflow Management Studied relationship between asynchronous data placement services and workflow management systems Workflow system can provide hints r.e. grouping of files, expected order of access, dependencies, etc. Contrasts with many existing workflow systems Explicitly stage data onto computational nodes before execution Some explicit data staging may still be required Data placement has potential to Significantly reduce need for on-demand data staging Improve workflow execution time Experimental evaluation demonstrates that good placement can significantly improve workflow execution performance Data Placement for Scientific Applications in Distributed Environments, Ann Chervenak, Ewa Deelman, Miron Livny, MeiHui Su, Rob Schuler, Shishir Bharathi, Gaurang Mehta, Karan Vahi, in Proceedings of Grid 2007 Conference, Austin, TX, September 2007. Approach: Combine Pegasus Workflow
Management with Globus Data Replication Service Compute Cluster Storage Elements Jobs Data Transfer Workflow Tasks Workflow Planner: Pegasus Staging Request Setup Transfers Data Placement Service: Globus
DRS Replication occurs when Replica Placement I want replica X at sites A, B, and C I want N replicas of each file I want replicas near my compute clusters Replica Repair Due to replica failure: lost or corrupted But it can be hard to tell the difference between permanent and temporary failure!
Examples of Placement Policies Topology-Aware Placement 1. Put Data client 2. Replicate to 2nd Local Site 3. Replicate to Remote Site Site 1 Site 2 Site 3 The Topology Aware policy is a type of N-copy policy that (in this 3-copy example) ensures that replicas are distributed within and between sites Publish/Subscribe Placement client 1.a. Publish Data XYZ 1.c. Subscribe XYZ and QRS client 1.b. Publish Data QRS Site 1 2. Query replica name service and replicate data sets
client Site 2 Site 3 The Publish/Subscribe policy is a query-based policy that identifies desired replicas based on a query and replicates them to the desired site Reactive vs. Proactive Replication Reactive Replication When a replica failure occurs, replicate Difficult to tell the difference between a permanent replica failure and a temporary loss e.g., temporary network partition Proactive replication Continually replicate files beyond the
minimum required Avoid bursts of network traffic to repair failures; limit bandwidth for repairs Need creation rate >= failure rate
