When data management meets cosmology and astrophysics: some lessons learned from the Petasky project Farouk Toumani LIMOS, CNRS, Blaise Pascal University Clermont-Ferrand, France ● Expérience de cosmologie de 4ème génération : ● Télescope de 8,4 m ● Cerro Pachon (Chili) ● Astronomie très grand champ : caméra 9,6□ ● ● Tout le ciel visible en 6 bandes optiques (20000 □) Poses de 15 s, 1 visite / 3 jours Journées de l’interdisciplinarité, 10-11 Décembre, 2014, Paris, France 10 ans, 60 Pbytes de données ● 15/03/12 Emmanuel Gangler – Réunion LIMOS Petasky http://com.isima.fr/Petasky Mastodons(program!of!the!Interdisciplinary!Mission!of!CNRS! • INS2I LIMOS (UMR CNRS 6158, Clermont-Ferrand) ✦ LIRIS (UMR CNRS 5205, Lyon) ✦ LABRI (UMR CNRS 5800, Bordeaux) ✦ LIF (UMR CNRS 7279, Marseille) ✦ LIRMM (UMR CNRS 5506, Montpellier) IN2P3 ✦ LPC (UMR CNRS 6533, Clermont-Ferrand) ✦ APC (UMR CNRS 7164, Paris) ✦ LAL (UMR CNRS 8607, Paris) ✦ Centre de Calcul de l’IN2P3/CNRS (CC-IN2P3) INSU ✦ LAM (UMR CNRS 7326, Marseille) ✦ • • Petasky: scientific challenges • Management of scientific data in the fields of cosmology and astrophysics ➡ Large amount of data ➡ Complex data (e.g., images, uncertainty, multi-scales...) ➡ Heterogeneous formats ➡ Various and complex processing (images analysis, reconstruction of trajectories, ad-hoc queries and processings, …) • Scientific challenges ➡ Scalability ➡ Data integration ➡ Data analysis ➡ Visualisation • Application context : LSST project Science in an exponential world The availability of very large amounts of data and the ability to efficiently process them is changing the way we do science • Science paradigms1 1. Empirical description of natural phenomena 2. Theoretical science: models and generalization 3. Computational science: simulation of complexe phenomena to validate theories 4. Data Intensive science : collecting and analyzing large amount of data 1Jim Gray, eScience Talk at NRC-CSTB meeting Mountain View CA, 11 January 2007. From Astronomy to astroinformatics • Modern digital detectors, CCDs, • Early use of scientific computing, numeric simulations, .. ➡ Antikythera mechanism, between 150 to 100 BC ➡ Supernovae Cosmology Project, 1986 ➡ 1024x1024 CCD camera, 2 megabytes every five minutes International Virtual Observatory Alliance (IVOA) - Web of astronomical data ➡ Sloan Digital Sky Survey (SDSS) ➡ GAIA, launched in 12/2013 and started the scientific observations in 7/2014 • A culture of sharing data ➡ Data with non-commercial value (more open than healthcare or biomedical science field) From Astronomy to astroinformatics • Modern digital detectors, CCDs, • Early use of scientific computing, numeric simulations, .. ➡ Antikythera mechanism, between 150 to 100 BC ➡ Supernovae Cosmology Project, 1986 ➡ 1024x1024 CCD camera, 2 megabytes every five minutes International Virtual Observatory Alliance (IVOA) - « … a mechanical computer used for calculating lunar, solar and stellar calendars » Web of astronomical data ➡ Sloan Digital Sky Survey (SDSS) ➡ GAIA, launched in 12/2013 and started the scientific observations in 7/2014 • A culture of sharing data ➡ Data with non-commercial value (more open than healthcare or biomedical science field) From Astronomy to astroinformatics • Modern digital detectors, CCDs, • Early use of scientific computing, numeric simulations, .. ➡ Antikythera mechanism, between 150 to 100 BC ➡ Supernovae Cosmology Project, 1986 ➡ 1024x1024 CCD camera, 2 megabytes every five minutes International Virtual Observatory Alliance (IVOA) - « … a mechanical computer used for calculating lunar, solar and stellar calendars » Web of astronomical data ➡ Sloan Digital Sky Survey (SDSS) ➡ GAIA, launched in 12/2013 and started the scientific observations in 7/2014 • A culture of sharing data ➡ Data with non-commercial value (more open than healthcare or biomedical science field) From Astronomy to astroinformatics • Modern digital detectors, CCDs, • Early use of scientific computing, numeric simulations, .. ➡ Antikythera mechanism, between 150 to 100 BC ➡ Supernovae Cosmology Project, 1986 - 1024x1024 CCD camera, 2 megabytes every five minutes « … a mechanical computer used for calculating lunar, solar and stellar calendars » ➡ International Virtual Observatory Alliance (IVOA) SDSS - Web of astronomical data ➡ Sloan Digital Sky Survey (SDSS) ➡ GAIA, launched in 12/2013 and started the scientific observations working in in 2000 • Started 7/2014 • In 2010, a total archive of 140 TB • 2.5 m Telescope, 54 CCD imager • A culture of sharing data ➡ Data with non-commercial value (more open than healthcare or biomedical science field) From Astronomy to astroinformatics • Modern digital detectors, CCDs, • Early use of scientific computing, numeric simulations, .. ➡ Antikythera mechanism, between 150 to 100 BC ➡ Supernovae Cosmology Project, 1986 - 1024x1024 CCD camera, 2 megabytes every five minutes « … a mechanical computer used for calculating lunar, solar and stellar calendars » ➡ International Virtual Observatory Alliance (IVOA) SDSS - Web of astronomical data ➡ Sloan Digital Sky Survey (SDSS) ➡ GAIA, launched in 12/2013 and started the scientific observations working in in 2000 • Started 7/2014 • In 2010, a total archive of 140 TB • 2.5 m Telescope, 54 CCD imager • A culture of sharing data ➡ Data with non-commercial value (more open than healthcare or biomedical science field) From Astronomy to astroinformatics • Modern digital detectors, CCDs, • Early use of scientific computing, numeric simulations, .. ➡ Antikythera mechanism, between 150 to 100 BC ➡ Supernovae Cosmology Project, 1986 - 1024x1024 CCD camera, 2 megabytes every five minutes « … a mechanical computer used for calculating lunar, solar and stellar calendars » ➡ International Virtual Observatory Alliance (IVOA) SDSS - Web of astronomical data ➡ Sloan Digital Sky Survey (SDSS) ➡ GAIA, launched in 12/2013 and started the scientific observations working in in 2000 • Started 7/2014 • In 2010, a total archive of 140 TB • 2.5 m Telescope, 54 CCD imager • A culture of sharing data ➡ Data with non-commercial value (more open than healthcare or biomedical science field) From Astronomy to astroinformatics • Modern digital detectors, CCDs, • Early use of scientific computing, numeric simulations, .. ➡ Antikythera mechanism, between 150 to 100 BC ➡ Supernovae Cosmology Project, 1986 - 1024x1024 CCD camera, 2 megabytes every five minutes « … a mechanical computer used for calculating lunar, solar and stellar calendars » ➡ International Virtual Observatory Alliance (IVOA) SDSS - Web of astronomical data ➡ Sloan Digital Sky Survey (SDSS) ➡ GAIA, launched in 12/2013 and started the scientific observations working in in 2000 • Started 7/2014 • In 2010, a total archive of 140 TB • 2.5 m Telescope, 54 CCD imager • A culture of sharing data ➡ Data with non-commercial value (more open than healthcare or biomedical science field) How much bytes… 10005 1015 péta 10004 1012 téra 10003 109 giga 10002 106 méga 10001 103 kilo How much bytes… 10005 1015 péta 10004 1012 téra A single text character giga A typewritten page A high-resolution photograph méga The complete works of Shakespeare 1 byte 2 kilobyte s 2 megabytes 5 megabytes A minute of high-fidelity sound kilo A pickup truck filled with books The contents of a DVD A collection of the works of Beethoven 50,000 trees made into paper and printed The print collections of the U.S. Library of Congress All U.S. academic research libraries 10 megabytes 1 gigabyte GB ) 17 gigabyte s 20 gigabytes 1 terabyte ( TB ) 10 terabytes 2 petabytes 10003 109 10002 106 10001 103 All hard disk capacity developed in 1995 http://searchstorage.techtarget.com/definition/How-many-bytes-for 20 petabytes Sizes of the astronomical datasets PB TB GB MB KB 1980 1990 2000 2010 Sizes of the astronomical datasets PB TB GB MB KB 1980 1990 2000 2010 Sizes of the astronomical datasets PB TB GB MB KB 1980 1990 2000 2010 Sizes of the astronomical datasets PB TB GB MB KB 1980 1990 2000 2010 Sizes of the astronomical datasets PB TB GB MB KB 1980 1990 2000 2010 Sizes of the astronomical datasets PB TB GB MB KB 1980 1990 2000 2010 E-science evolution E-science evolution E-science evolution E-science evolution Homo FTP-GREPus E-science evolution Homo FTP-GREPus E-science evolution Homo FTP-GREPus In 20041 FTP/GREP 1GB in a minute FTP/GREP 1TB in 2 days FTP/GREP 1PB in 3 years 1Where The Rubber Meets the Sky Giving Access to Science Data, Jim Gray and Alex Szalay E-science evolution Homo FTP-GREPus In 20041 FTP/GREP 1GB in a minute FTP/GREP 1TB in 2 days FTP/GREP 1PB in 3 years 1Where The Rubber Meets the Sky Giving Access to Science Data, Jim Gray and Alex Szalay E-science evolution Homo FTP-GREPus In 20041 FTP/GREP 1GB in a minute FTP/GREP 1TB in 2 days FTP/GREP 1PB in 3 years Homo Numericus 1Where The Rubber Meets the Sky Giving Access to Science Data, Jim Gray and Alex Szalay E-science evolution Homo FTP-GREPus In 20041 FTP/GREP 1GB in a minute FTP/GREP 1TB in 2 days FTP/GREP 1PB in 3 years Homo Numericus 1Where The Rubber Meets the Sky Giving Access to Science Data, Jim Gray and Alex Szalay E-science evolution Homo FTP-GREPus In 20041 FTP/GREP 1GB in a minute FTP/GREP 1TB in 2 days FTP/GREP 1PB in 3 years Homo Numericus 1Where The Rubber Meets the Sky Giving Access to Science Data, Jim Gray and Alex Szalay Grid computing Cloud computing Virtualization MapReduce New hardware NoSQL … Data-driven discovery in Astrophysics Telescopes Observatories Data-driven discovery in Astrophysics Telescopes Observatories Digitized data Data-driven discovery in Astrophysics Telescopes Observatories Digitized data Processing pipeline Data-driven discovery in Astrophysics Telescopes Observatories Digitized data Processing pipeline Information/ knowledge Data-driven discovery in Astrophysics Composants d’un Composants SGBD d’un Composants SGBD d’un SGBD Telescopes Observatories User/Application User/Application Composants d’un SGBD DBA Composants DBA d’un SGBD Digitized Transaction commands Queries, updates Transaction commands Queries, updates data Query Transaction Metadata, compiler manager statistics Transaction commands Query Transaction DDL Transaction Execution Logging and manager engine recovery Index, file and record requests Logging and Index/file/record Log recovery manager pages Data, BufferBuffer manager Data, metadata, indexes Storage Storage Buffer manager Storage Log pages Page commands DDL commands Transaction DDL DDL Query Execution Logging Concurrency and compiler Metadata, compiler engine recoverycontrolstatistics Processing Query plan pipeline Index/file/record Execution Concurrency Log control manager engine Lock tableData, pages Data, Index, metadata, file and record Page indexes requests commands metadata, ge indexes ands DBA DDL commands Metadata, compiler Metadata, manager compiler manager compiler DBA User/Application statistics statistics Metadata Transaction commands Query DDL commands Queries, plan updates Query plan Metadata, statistics ex, file record quests Transaction commands DDL commands Transaction Logging Concurrency and manager recoverycontrol Metadata Metadata DDL compiler DBA DDL Concurrency compiler control Metadata Logging and Log recoveryLock table pages Concurrency control Lock table metadata, indexes Information/ BufferBuffer Index/file/record knowledge manager manager Lock tableData, Buffer metadata, indexes Storage Buffer Storage manager manager Storage Buffer Dedicated Storage manager archives Storage Metadata DDL commands Log pages Lock table Data-driven discovery in Astrophysics Composants d’un Composants SGBD d’un Composants SGBD d’un SGBD Telescopes Observatories User/Application User/Application Composants d’un SGBD DBA Composants DBA d’un SGBD Digitized Transaction commands Queries, updates Transaction commands Queries, updates data Query Transaction Metadata, compiler manager statistics Transaction commands Query Transaction DDL Transaction Execution Logging and manager engine recovery Index, file and record requests Logging and Index/file/record Log recovery manager pages Data, BufferBuffer manager Data, metadata, indexes Storage Storage Buffer manager Storage Log pages Page commands DDL commands Transaction DDL DDL Query Execution Logging Concurrency and compiler Metadata, compiler engine recoverycontrolstatistics Processing Query plan pipeline Index/file/record Execution Concurrency Log control manager engine Lock tableData, pages Data, Index, metadata, file and record Page indexes requests commands metadata, ge indexes ands DBA DDL commands Metadata, compiler Metadata, manager compiler manager compiler DBA User/Application statistics statistics Metadata Transaction commands Query DDL commands Queries, plan updates Query plan Metadata, statistics ex, file record quests Transaction commands DDL commands Transaction Logging Concurrency and manager recoverycontrol Metadata Metadata DDL compiler DBA DDL Concurrency compiler control Metadata Logging and Log recoveryLock table pages Concurrency control Lock table metadata, indexes Information/ BufferBuffer Index/file/record knowledge manager manager Lock tableData, Metadata DDL commands Public Buffer Log pages Lock table metadata, indexes Storage Buffer Storage manager manager Storage Buffer Dedicated Storage manager archives Storage Education Science Data-driven discovery in Astrophysics Composants d’un Composants SGBD d’un Composants SGBD d’un SGBD Telescopes Observatories User/Application User/Application Composants d’un SGBD DBA Composants DBA d’un SGBD Digitized Transaction commands Queries, updates Transaction commands Queries, updates data Query Transaction Metadata, compiler manager statistics Transaction commands Query Transaction DDL Transaction Execution Logging and manager engine recovery Index, file and record requests Logging and Index/file/record Log recovery manager pages Data, BufferBuffer manager Data, metadata, indexes Storage Storage Buffer manager Storage Log pages Page commands DDL commands Transaction DDL DDL Query Execution Logging Concurrency and compiler Metadata, compiler engine recoverycontrolstatistics Processing Query plan pipeline Index/file/record Execution Concurrency Log control manager engine Lock tableData, pages Data, Index, metadata, file and record Page indexes requests commands metadata, ge indexes ands DBA DDL commands Metadata, compiler Metadata, manager compiler manager compiler DBA User/Application statistics statistics Metadata Transaction commands Query DDL commands Queries, plan updates Query plan Metadata, statistics ex, file record quests Transaction commands DDL commands Transaction Logging Concurrency and manager recoverycontrol Metadata Metadata DDL compiler DBA DDL Concurrency compiler control Metadata Logging and Log recoveryLock table pages Concurrency control Lock table metadata, indexes Information/ BufferBuffer Index/file/record knowledge manager manager Lock tableData, Metadata DDL commands Public Buffer Log pages Lock table metadata, indexes Storage Buffer Storage manager manager Storage Buffer Dedicated Storage manager archives Storage Education Science Astrometric pipeline Outputs object positions Astrometric pipeline Outputs object positions Spectroscopic pipeline redshifts, classification of objects, … ● Expérience de cosmologie de 4ème génération : The LSST project Télescope de 8,4 m Large Synoptic Survey telescope Cerro Pachon (Chili) ● ● Astronomie très grand champ : □ caméra 9,6 Telescope of 8.4 m ● A new window over the sky: ● ● ● 15/03/12 Tout le ciel visible en 6 bandes optiques (20000 □) Poses de 15 s, 1 visite / 3 jours 10 ans, 60 Pbytes de données Emmanuel Gangler – Réunion LIMOS The LSST Organisation projet : PetaSky Non-proft corporation ● US : 33 partenaires ; 670 M$ ● Chili : site ● France : IN2P3 (~15 M€) Telescope Caméra Data Management Outreach « The data volumes […] of LSST are so large that the limitation on our ability to do science isn't the ability to collect the data, it's the ability to understand […] the data » Andrew Conolly (U. Washington) “How do you turn petabytes of data into scientific knowledge?” Kirk Borne (George Mason U.) Data management challenges in LSST “How much the (LSST) project will tell us about our solar system, the dark energy problem and more, will depend on how well we can process the information the telescope and its camera send back to us - an estimated sum of around ten petabytes of data per year.” (Mari Silbey, Space: the big data frontier, http://www.smartplanet.com/blog/thinking-tech/space-the-big-data-frontier/12180) Data management challenges in LSST “How much the (LSST) project will tell us about our solar system, the dark energy problem and more, will depend on how well we can process the information the telescope and its camera send back to us - an estimated sum of around ten petabytes of data per year.” (Mari Silbey, Space: the big data frontier, http://www.smartplanet.com/blog/thinking-tech/space-the-big-data-frontier/12180) “Plans for sharing the data from LSST with the public are as ambitious as the telescope itself” Anyone with a computer will be able to fly through the Universe, zooming past objects a hundred million times fainter than can be observed with the unaided eye. The LSST project will provide analysis tools to enable both students and the public to participate in the process of scientific discovery. LSST data scales Tens of thousands of billions of photometric observations over tens of billions of objects LSST data scales Tens of thousands of billions of photometric observations over tens of billions of objects • 1-10 Millions events per night, 3 billions of sources • 16 TB each 8 hours at a rate of 540 MB/second Images • • - 12,8 GB/39 per second (data rate of 330 MB/s) - 100 PB final archives of images Transients - • • 1-3 PB, Relation with 100 attributes/5000 billions of tuples Objects catalog : - Relation with 500 attributes, 40 Billions of tuples - 100-200 TB Estimation for the end of the project • 400 000 Billions of tuples (different versions of data in addition to replication), • Raw data: 60 PB • Catalog database: 15 PB • Total volume after processing : several hundred of PB LSST data scales Tens of thousands of billions of photometric observations over tens of billions of objects • 1-10 Millions events per night, 3 billions of sources • 16 TB each 8 hours at a rate of 540 MB/second Images • • - 12,8 GB/39 per second (data rate of 330 MB/s) - 100 PB final archives of images Transients - • • Highest data rate in current astronomical surveys: 4.3 MB/s Sloan Digital Sky Survey (SDSS) 1-3 PB, Relation with 100 attributes/5000 billions of tuples Objects catalog : - Relation with 500 attributes, 40 Billions of tuples - 100-200 TB Estimation for the end of the project • 400 000 Billions of tuples (different versions of data in addition to replication), • Raw data: 60 PB • Catalog database: 15 PB • Total volume after processing : several hundred of PB LSST data scales LSST year 1 LSST year 10 Raw data 6 PB 60 PB Archive 19 PB 270 PB Disk (DAC) 16 PB 90 PB DB (baseline) 0,5 PB 5 PB Moore equivalent 2014 1.2 TB 12 TB LSST data scales LSST year 1 Table& LSST year 10 Raw data Archive Disk (DAC) 6 PB Object' 60 PB 109'TB' Moore equivalent 2014 #tuples& #a/ributes& 38'B' 470' 19 PB Moving'Object' 270 PB 5'GB' 6'M' 100' 16 PB Source' 5'T' 125' 32'T' 7' 0,5 PB Difference'Image' 5 PB 71'TB' Source' 200'B' 65' 1.2 TB CCD'Exposure' 12 TB 0.6'TB' 17'B' 45' 90 PB 3.6'PB' Forced'Source' DB (baseline) Size& 1.1'PB' LSST data scales A change of scale from TB to PB Data volumes & rates are unprecedented in astronomy Relative Etendue (= A ) 320 280 20000 Estimated Nightly Data Volume Etendue (m2 deg2) 240 15000 200 160 GB 10000 120 5000 80 All facilities assumed operating100% in one survey 0 40 0 LSST PS4 PS1 Subaru CFHT SDSS MMT Steven M. Kahn SLUO Meeting Presentation June 7, 2007 DES 4m VST Raw Catalog LSST Pan-STARRS 4 VISTA IR 4 SDSS LSST will make tens of trillions photometric observations of tens of billions of objects Queries per difficulty level Supported queries • Retrieve any type of information about a single object (identified by a given objectId), including full time series. SELECT * FROM Object JOIN Source USING (objectId) WHERE objectId = 293848594; Few seconds • Retrieve any type of information about a group of objects in a small area of sky, including neighborhood-type queries. SELECT * FROM Object WHERE qserv_areaSpec_circle(1.0, 35.0, 5.0/60) ≃1 hour • Analysing light curves across large area. SELECT O.objectId, myFunction(S.taiMidPoint, S.psfFlux) FROM Object AS O JOIN Source AS S USING (objectId) WHERE O.varProb > 0.75 GROUP BY O.objectId; ≃1 day (24h) • Analysing light curves of faint objects across large area. SELECT O.objectId, myFunction(V.taiMidPoint, FS.flux) FROM Object AS O JOIN ForcedSource AS FS ON (O.objectId = FS.objectId) JOIN Visit AS V ON (FS.visitId = V.visitId); ≃1 week ! ! Queries per difficulty level Expensive/impossible queries • Expensive queries – Find objects far away from other objects (for a large number of objects). Question: what is the largest distance we should plan to support for distance-based queries involving (a) small number of objects, (b) all objects on the sky? – Sliding window queries: Find all 5 arcmin x 5 arcmin regions with an object density higher than rho • Impossible queries – Large size results • Select all pairs of stars within 1 arc min of each other in the Milky Way region. –Expensive or hidden computation (e.g., Join) • Near neighbor query on the Source or ForcedSource table • Joining large tables between different LSST data releases • Time series analysis of every object • Cross-match with very large external catalog (e.g. LSST with SKA) • Any non-spatial join on the entire catalog (Object, Source, ForcedSource) • Join of Source with ForcedSource Examples of User Defined Functions • • • q3c_ang2ipix(ra, dec) -- returns the ipix value at ra and dec • q3c_ellipse_join(ra1, dec1, ra2, dec2, major, ratio, pa) -- like q3c_join, except (ra1, dec1) have to be within an ellipse with major axis major, the axis ratio ratio and the position angle pa (from north through east) • q3c_radial_query(ra, dec, center_ra, center_dec, radius) -- returns true if ra, dec is within radius degrees of center_ra, center_dec. This is the main function for cone searches. function should be used if when the index on q3c_ang2ipix(ra,dec) is created) • q3c_ellipse_query(ra, dec, center_ra, center_dec, maj_ax, axis_ratio, PA ) -- returns true if ra, dec is within the ellipse from center_ra, center_dec. The ellipse is specified by major axis, axis ratio and positional angle. function should be used if when the index on q3c_ang2ipix(ra,dec) is created) • • • q3c_poly_query(ra, dec, poly) -- returns true if ra, dec is within the postgresql polygon poly. • q3c_ipixcenter(ra, dec, bits) -- the function returning the ipix value of the pixel center of certain depth covering the specified (ra,dec) q3c_dist(ra1, dec1, ra2, dec2) -- returns the distance in degrees between (ra1,dec1) and (ra2,dec2) q3c_join(ra1, dec1, ra2, dec2, radius) -- returns true if (ra1, dec1) is within radius spherical distance of (ra2, dec2). It should be used when the index on q3c_ang2ipix(ra2,dec2) is created. q3c_ipix2ang(ipix) -- returns a 2-array of (ra,dec) corresponding to ipix. q3c_pixarea(ipix, bits) -- returns the area corresponding to ipix at level bits (1 is smallest, 30 is the cube face) in steradians. Main thesis of this talk How to form a new generation of scientists capable to exploit the new technologies to pursue science goals at an unprecedent scale? G.Longo 1 1Talk, workshop « New challenges in astro- and environmental informatics in the Big Data era », May, 2014, Szombathely, Hungary Main thesis of this talk How to form a new generation of scientists capable to exploit the new technologies to pursue science goals at an unprecedent scale? G.Longo 1 It should not be up to the scientists but to the technology (data management system) to overcome the computing barriers between them and the data 1Talk, workshop « New challenges in astro- and environmental informatics in the Big Data era », May, 2014, Szombathely, Hungary Main thesis of this talk How to form a new generation of scientists capable to exploit the new technologies to pursue science goals at an unprecedent scale? G.Longo 1 It should not be up to the scientists but to the technology (dataAnalysts management system) to instead of Hackers overcome the computing barriers between them and the data 1Talk, workshop « New challenges in astro- and environmental informatics in the Big Data era », May, 2014, Szombathely, Hungary Main thesis of this talk How to form a new generation of scientists capable to exploit the new technologies to pursue science goals at an unprecedent scale? G.Longo 1 It should not be up to the scientists but to the technology (dataAnalysts management system) to instead of Hackers overcome the computing barriers between them and the data 1Talk, workshop « New challenges in astro- and environmental informatics in the Big Data era », May, 2014, Szombathely, Hungary Main thesis of this talk This approach has been very successful in How to form a new generation of scientists capable to business domain but intogeneral not sogoals successful exploit the new technologies pursue science at an in the scientific domain unprecedent scale? G.Longo 1 It should not be up to the scientists but to the technology (dataAnalysts management system) to instead of Hackers overcome the computing barriers between them and the data 1Talk, workshop « New challenges in astro- and environmental informatics in the Big Data era », May, 2014, Szombathely, Hungary What makes DB technology successful in business domain? • Abstractions ✓ Relation instead of files, blocks, tablespaces, segments, extents, access path ✓ Relational algebra instead of algorithms • Declarative query language ✓ Express what you want not how to get it • Optimization ✓ Rather naive techniques but enough for the business world Example: query compilation Select Name From Employé E, Département D, Projet.P Where E.empno=D.manager and E.empno=P.mgr Example: query compilation Exemple (cont.) Select Name From Employé E, Département D, Projet.P Plan d’exécution Where E.empno=D.manager and E.empno=P.mgr Index Nested loop (A.x=C.x) Merge-join (A.x=B.x Sort Table-scan A Table-scan Employé Index scanProjet C Index-scan Sort Table-scan B Table-scan Département Petasky: data management challenge Techniques to build an efficient and easy to use data access system at a reasonable cost • • • Specialized Hardware Programming Ad-hoc optimization • • • Commodity machines Querying Generic system Space of solutions and associated challenges ! Clearly beyond the capacities of centralized systems • • Distributed and parallel systems ✓ Data distribution ✓ Computation distribution ✓ Failure resilience Storage model ✓ row store vs. column store ✓ (sophisticated) Indexes • Benefit from modern hardware • Complexity theory and cost models ✓ Standards measures: I/O, data transfer, .. ✓ Cost of coordination Explored approaches • Big data approaches ✓ ✓ Distributed and parallel systems ➡ MapReduce-like approaches (shared nothing architecture) ➡ Parallel DBMS (shared all thing architecture) ➡ Spatial partitioning (QSERV for LSST) Column store DBMSs ➡ • Vertica, MonetDB, ... Data integration to the rescue ➡ Declarative approach ● Agrégation des résultats ● Le temps de calcul assage à grande échelle (au CC IN2P3) : œuds physiques, 120 GB disque, 16 GB RAM es, 3000 chunks → 50 GB / nœud lélisation du déploiement : espace disque = taille DB x2 23/1/14 Emmanuel Gangler – Workshop Mastodons des surcoûts (tables en cache) Limitations • ités de performance : Limited queries ✓ - distance ~1 arcmin → 300 noeuds - non spatial joins munications, multi-threading Non-partitionable aggregates réponse ~proportionnel - de ✓ User defined functions mbre de nœuds ✓ Load balancing ✓ Ad-hoc query rewriting 13/40 Mediation-based approach (cont.) Data integration at the rescue User query Answers Semantic Mappings . . . View on Source 1 Answers View on Source 2 View on Source 3 View on Source N Query Wrapper 1 Answers Wrapper 2 Wrapper 3 Wrapper N Query Documents Source 1 Source 2 Source 3 Source N uted on clusters of several nodes and both their inputs and outputs are files ibuted file system (e.g.,Mapreduce Hadoop Distributed File System (HDFS)). approaches a new programming model intended to facilitate the 1 development MapReduce of Execution Overview scalable parallel computations on large server clusters Mapreduce approaches Hadoop! Hadoop++! HAIL! HadoopDB! Hive! Query!language procedural Procedural Procedural DeclaraGve DeclaraGve Simple!indexes Not!supported supported Supported supported Supported MulG!indexes Not!supported Not!supported supported Supported Supported Complex!indexes Not!supported Not!supported Not!supported supported Supported Storage! System HDFS HAIL Classical!DBMS HDFS Manual Manual AutomaGc Manual [Dean!et!al,!04] Index!choice HDFS O [Di3rich!et!al,!10] [Di%rich!et!al,!12] [Abouzeid!et!al,!09] [Thusoo!et!al,!10] Data Mapreduce loading (1/2) approaches Minutes" 2000" 1500" Hadoop! Hadoop++! HAIL! HadoopDB! Hive! procedural Procedural Procedural DeclaraGve DeclaraGve supported Supported Not!supported supported Supported Difference"(Hive">>"HadoopDB):"300"%"for" 25"machines"and"200"%"for"50"machines" • Hive:" Supported [Di3rich!et!al,!10] [Dean!et!al,!04] 1000" 500" Query!language 0" 250"GB" 500"GB" Simple!indexes 1"TB" 25"machine" Not!supported Hive"(HDFS)" global"hash" 250"GB" supported local"hash" [Di%rich!et!al,!12] 500"GB" [Abouzeid!et!al,!09] 1"TB" 50"machines" Supported tuning" loading.4me.for.1.TB9.50.machines. MulG!indexes Not!supported • 33%. 39%. Complex!indexes Not!supported 15%. Storage! System HDFS" global"hash" Index!choice 13%. Not!supported • Not!supported 2X"of"Ime"for"2X"of"data"volume"" • 25"!"50"machines:"same"Ime" • HadoopDB:" HDFS local"hash" tuning" supported Supported HDFS HAIL Classical!DBMS HDFS Manual Manual AutomaGc Manual • 2X"of"data:"90"%K120"%""supplementary"" • 25"!50"machines:"25"%""gain"" O [Thusoo!et!al,!10] Data Mapreduce loading (1/2) approaches Data loading (2/2) 1500" Hadoop! Hadoop++! HAIL! 6%$ HadoopDB! Hive! procedural Procedural Procedural DeclaraGve DeclaraGve supported Supported [Di3rich!et!al,!10] [Dean!et!al,!04] 30%$ 1000" 500" Query!language 250"GB" 500"GB" 1"TB" 25"machine" Not!supported Simple!indexes Load%&me% 250"GB" supported indexing% Hive"(HDFS)" global"hash" local"hash" Indexing$for$1$TB$3$50$machines$$ Hive$ loading.4me.for.1.TB9.50.machines. MulG!indexes Not!supported • 200% 33%. 150% 39%. 15%. Storage! System global"hash" Index!choice 100% 13%. 50% 0% [Abouzeid!et!al,!09] 94%$ Load%&me% [Thusoo!et!al,!10] indexing% Indexing$for$1$TB$3$50$machines$ HadoopDB$ Not!supported supported Supported Difference"(Hive">>"HadoopDB):"300"%"for" • Hive:% 25"machines"and"200"%"for"50"machines" objec&d% O ra% Not!supported • Not!supported 2X"of"Ime"for"2X"of"data"volume"" • 25"!"50"machines:"same"Ime" decl% Supported • 25%!%50%machines:%gain%of%15%% supported Supported HAIL Classical!DBMS HDFS Manual AutomaGc Manual • 2X"of"data:"90"%K120"%""supplementary"" 500%GB% • 25"!50"machines:"25"%""gain"" 1TB% 2%TB% 250GB% Sourceid% tuning" HDFS tuning" 1"TB" 50"machines" Supported • HadoopDB:" HDFS local"hash" 500"GB" • Hive:" Complex!indexes Not!supported HDFS" [Di%rich!et!al,!12] 70%$ 0" Index%size%(GB)% Minutes" 2000" scienceccdexposure% Manual Data Mapreduce loading (1/2) approaches Data loading (2/2) 1500" Hadoop! Hadoop++! HAIL! 6%$ HadoopDB! Hive! procedural Procedural Procedural DeclaraGve DeclaraGve supported Supported Query processing Hive vs. HadoopDB (2/6) [Di3rich!et!al,!10] [Dean!et!al,!04] 30%$ 1000" 500" Query!language 500"GB" 1"TB" 2000" 250"GB" 1800" 25"machine" Not!supported 1600" Load%&me% supported indexing% Time"(seconds)" 1400" Hive"(HDFS)" global"hash" local"hash" Indexing$for$1$TB$3$50$machines$$ 1200" Hive$ 1000" loading.4me.for.1.TB9.50.machines. MulG!indexes 800"Not!supported • 200% 600" 33%. 400" 39%. 150% 200" [Thusoo!et!al,!10] [Abouzeid!et!al,!09] global"hash" Storage! System Index!choice 100% Q1" 50% tuning" HDFS 160" Load%&me% 140" indexing% Indexing$for$1$TB$3$50$machines$ 120" HadoopDB$ 100" 0" Q2" Q3" Q4" HDFS objec&d% O ra% Supported Hive" HadoopDB" Not!supported Hive" Supported Hive" supported HadoopDB" Hive" HadoopDB" Hive" HadoopDB" • Not!supported 2X"of"Ime"for"2X"of"data"volume"" 500"GB" 1"TB" 250"GB" 500"GB" 1"TB" • 25"!"50"machines:"same"Ime" Q1" Q2" decl% scienceccdexposure% Manual Q3"(RA)" Q4" HAIL Classical!DBMS HDFS Manual AutomaGc Manual • 2X"of"data:"90"%K120"%""supplementary"" 500%GB% • 25"!50"machines:"25"%""gain"" 1TB% 2%TB% 250GB% Sourceid% 94%$ • 25%!%50%machines:%gain%of%15%% 20" • HadoopDB:" local"hash" 0% tuning" HadoopDB" 250"GB" 13%. 200" 50"machines" Supported180" • Hive:" 0" 1"TB" Not!supported supported80" Supported Difference"(Hive">>"HadoopDB):"300"%"for" • Hive:%60" 25"machines"and"200"%"for"50"machines" 40" HadoopDB" Hive" Complex!indexes Not!supported 15%. 500"GB" Time"(seconds)" 250"GB" Simple!indexes HDFS" [Di%rich!et!al,!12] 70%$ 0" Index%size%(GB)% Minutes" 2000" Physical storage Row store vs. column store Row$Oriented$Database$ • High cost of I/O • Analytical queries are expensive • Problem with null values Column$Oriented$Database$ • Cost of join Hybrid storage T T1 T2 Hybrid Database Hybrid storage T Column1( ORACLE( MonetDB( Hybrid( Q1( 00:05:00,00( 00:00:46.00( 00:06:00,00( Q2( 00:04:42,00( 00:00:00.12( 00:02:28,00( Q3( 00:04:15,00( 00:00:00.03(( 00:00:01,00( Q4( 00:04:04,00( Q5( 00:00:02,19( Q6( T1 T2 00:00:00.90( 00:03:40,00( 00:00:42.00( 00:05:06,00( 00:00:02,50( 00:00:03,00( 00:39:00,00( Q7( 00:00:18,14( 00:00:6.3( 00:00:18,00( Q8( 00:07:38,00( 00:00:39.2( 01:40:00,00( Q9( 00:01:39,00( 00:00:45.00( Q10( 00:04:03,00( (00:08:13,00( Hybrid Database 00:00:09,00( Learned lessons and research directions No one fits all • • • MapReduce-based algorithms can be useful to implement physical operators Hybrid system: row/column store Need for more research on ✓ Abstraction adequate to the scientific domain - Array data model (SCiDB)? ✓ Support of user defined functions ✓ Optimization techniques embedded in the data management system ✓ Scalability of information integration framework Scientists Data mining/learning/statistics Foundational approach to build scientific data management system Data management Visualisation Working cross disciplines … working cross cultures Scientists Data mining/learning/statistics Foundational approach to build scientific data management system Data management Visualisation Merci
© Copyright 2024 Paperzz