Overview

1. Introduction to Observational DataBase (ODB) sami.saarinen@ecmwf.intpaul.burton@ecmwf.int25-Apr-2007

2. Overview • Introduction to ODB • Creating a simple database • Use of simulobs2odb –program • Visualizing data using basic odbviewer • More complex databases • ODB within IFS/4DVAR-system • Manipulating ODB data from Fortran90 • Few tools: odbsql, odbdiff, odbcompress, odbdup, odb2netcdf • ODBTk : A GUI-based ODB visualisation toolkit • A separate presentation & demo by Paul Burton

3. Introduction to ODB • ODB is a tailor made (hierarchical) database software developed at ECMWF to manage very large observational data volumes through the ECMWF IFS/4DVAR-system on highly parallel supercomputer systems • ODB also enables flexible post-processing of observational data even on a desktop computer • ODB software is written in C and Fortran-90 languages and is available virtually on any Unix-systems (and now also for Windows/CYGWIN) • The software can be installed from source code (“tar-ball“) normally in a less than an hour

4. Snapshot of AIRS channel#1873 brightness-T

5. A snapshot of SATOB/AMV-winds

6. One month of averaged Br-T : HIRS,channel#4

7. … Introduction to ODB • An observational database usually contains following items: • Observation identification, position and time coordinates • Observation value, pressure levels, channel numbers • Various quality control flags • Obs. departures from background and analysis fields • Satellite specific information • Other closely related information • All information can be accessed via ODB/SQL language and Fortran90 interface • Also a direct (read-only) access to ODB-data is now available • no programming effort to “scan” ODB-data

8. Basic components of ODB • ODB/SQL-language • Data Definition Language: To describe what data items belong to database, what are their data types and how they are related (if any) to each other • Data Query Language: To query and return a subset of data which satisfies certain user specified conditions. This is the key feature of the ODB software !! • Fortran90 interface layer • Data manipulation : create, update & remove data • Execute ODB/SQL-queries and retrieve filtered data • To control MPI and OpenMP-parallelization

9. Creating a simple ODB database • We will create a very simple database using text files • The 3 text files describe • Data layout i.e. what data items will go into ODB • Location and time information of observations • Actual observation measurement information for each location at the given pressure levels • Feed these files into simulobs2odb-program • Discover the data values in database by using odbviewer

10. Data definition layout : MYDB.ddl CREATE TABLE hdr AS ( seqno pk1int, obstype pk1int, codetype pk1int, lat pk9real, lon pk9real, date yyyymmdd, time hhmmss, body@LINK, ); CREATE TABLE body AS ( entryno pk1int, varno pk1int, vertco_type pk1int, press pk9real, obsvalue pk9real, );

11. Input file#2 : hdr.txt #hdr obstype = 2 codetype = 141 seqno lat lon date time body.len 1 45 -15 20041101 000000 1

12. Input file#3 : body.txt #body entryno varno vertco_type press obsvalue 1 2 1 50000 251.0

13. Running simulobs2odb • Initialize ODB interactive environment : • use odb • Create database using the following simple command : • simulobs2odb –l MYDB –i hdr.txt –i body.txt • As a result of these commands, a small database called MYDB has been created and it contains one data pool with two tables hdr and body, which are linked (related) to each other via special @LINK data type • It is now easy to extend database by providing more data, or specifying more data items, or adding more tables, or all above at the same time

14. Visualizing with odbviewer • History: odbviewer was originally written to be used as a debugging tool for ODB software development • Linked with ECMWF graphics package MAGICS/MAGICS++ • Displays coverage plots • Also a textual report generator • Displays output of data queries • “Sensitive” to ODB/SQL-language : tries automatically produce both coverage plot and textual report for the user • Textual report itself can be invaluable source of information for further post-processing tasks • Making use of the new and more economical tool odbsql

15. Running odbviewer • Go to database directory • cd MYDB • Run • odbviewer –q ‘SELECT lat,lon,press,obsvalue\ FROM hdr, body \ WHERE obstype = 2’

16. odbviewer coverage plot Our observation !!

17. Some odbviewer options -h List of options (gimme some “help” !) -q ‘SQL-stmt’ Provide ODB/SQL-statement inline -v viewname/poolno Choose SQL name (& optionally pool number) -p “1-10,12,15” Choose from a subset of pools -R No radians-to-degrees conversion for (lat,lon) -r Enforce radians-to-degrees conversion -k Show (lat,lon) in degrees even if they were in radians in DB -c Clean start (i.e. recompile all) -e editor Choose preferred editor -e batch Run in batch mode (same as –e pipe) -N Do not produce a report at all -I Do not show plot immediately -P projection Change display projection -C file.cmap Supply a color map file -A plot_area Choose plotting area -F (en)Force to use the old style odbviewer over ‘odbsql’

18. More complex databases • In reality databases usually contain many more tables (>>5) than in the simple example earlier • Each table can contain 10—50 data columns • There can also be a sophisticated data hierarchy (see the next slide) to describe potentially quite complex relationships between tables • In order to provide a good parallelperformance on supercomputers, data tables are furthermore divided into data pools, which enables parallel I/O, too: • They behave like sub-databases within a database • Allows much bigger data sets than otherwise possible

19. Comprehensive data hierarchy

20. ECMA/ODB CCMA/ODB Output BUFRs ODB within IFS/4DVAR-system

21. AMSU-A data before screening

22. AMSU-A data after screening Under 10% left active !!

23. Typical ODB usage at ECMWF … • Database can be created interactively or in batch mode • We usually run our in-house BUFR2ODB in batch-mode • New observation types can also be fed in via text file • Complete database manipulation prefer using Fortran90-interface, but any read/only-database can also be accessed via rudimentary client-server –interface (C/C++) • Another possibility is to run the new tool – odbsql • No need to use of ODB/SQL compilation system • No need to write a single line of Fortran90 • The tool is under development

24. … Typical ODB usage at ECMWF • When database has been created, the application program queries data via precompiled ODB/SQL and places the result data (also known as view ) into a data matrix allocated by the user program • There can virtually be any number of active viewsat any given time. These can be updated and fed back to database • Due to ODB, the use of WMO BUFR has therefore been minimized at ECMWF in order to enable faster and more robust processing of observations

25. ECMWF BUFR to ODB conversion • ODBs at ECMWF are normally created by using bufr2odb • Enables MPI-parallel database creation  efficient • Allows retrospective inspection of Feedback BUFR data by converting it into ODB (slow & not all data in BUFR) • bufr2odb can also be used interactively, for example: bufr2odb –i bufr_input_file –I 1-20 –n 4 • The preceding example creates 4 pools of ECMA database from the given BUFR input file, but includes only BUFR subtypes from 1 to 20 (inclusive) • Feedback BUFR to ODB works similarly: fb2odb –i feedback_bufr_file –n 8 –u 2

26. Manipulating ODB from Fortran90 • Currently Fortran90 is the only way to fill an ODB database • simulobs2odb is also a Fortran90-program underneath • likewise odbviewer or practically any other ODB-tool • Also: to fetch and update data, Fortran90 is necessary • ODB Fortran90 interface layer offers a comprehensive set of functions to • Open & close database • Attach to & execute precompiled ODB/SQL queries • Load, update & store queried data

27. An example ODB program program main use odb_module implicit none integer(4) :: h, rc, nra, nrows, ncols, npools, j, jp real(8), allocatable :: x(:,:) npools = 0 h = ODB_open(‘MYDB’, ’OLD’, npools=npools) < data manipulation loop ; see next page > rc = ODB_close(h, save=.TRUE.) end program main

28. Data manipulation loop DO jp=1,npools ! Execute SQL, allocate space, get data into matrix rc = ODB_select(h,’sqlview’,nrows,ncols,poolno=jp) allocate(x(nrows,0:ncols)) rc = ODB_get(h,’sqlview’,x,nrows,ncols,poolno=jp) ! Update data, put back to DB, deallocate space call update(x,nrows,ncols) ! Not an ODB-routine rc = ODB_put(h,’sqlview’,x,nrows,ncols,poolno=jp) deallocate(x) rc = ODB_cancel(h,’sqlview’,poolno=jp) ! Use the following only with READONLY-databases ! rc = ODB_release(h,poolno=jp) ENDDO

29. Compile, link and run • use odb # once per session • (2) odbcomp MYDB.ddl # once only;often from file MYDB.sch • (3) odbcomp sqlview.sql # recompile only when changed • (4) odbf90 main.F90 update.F90 –lMYDB –o main.x # link • (5) ./main.x # run

30. ODB/SQL compilation system

31. odbsql • A new tool to access ODB data in read/only –mode • Does not generate C-code, but dives directly into data • Usually faster than generated C-code with exception of accessing large amounts of satellite data (investigated) • The tool is under active development right now • Usage: odbsql –q ‘SELECT column(s) FROM table(s) WHERE …’ \ –s starting_row –n number_of_rows_to_display \ [–X] [other_options]

32. ODB/SQL – examples (1) SET $t2m = 39; // Scalar parameters, whose values … SET $synop = 1; // … can be overridden in Fortran90 CREATE VIEW t2m AS SELECT an_depar, fg_depar, lat, lon, obsvalue FROM hdr, body WHERE obstype = $synop // Give me synops AND varno@body = $t2m // Give me 2 meter temperatures AND obsvalue is notNULL ; // Don’t want missing data

34. ODB/SQL – examples (2) SELECT count(*), avg(obsvalue), stdev(fg_depar) FROM hdr, body WHERE obstype = $synop && varno = $t2m AND obsvalue IS NOT NULL; // Observation count per (obstype,codetype)-pair : SELECTobstype, codetype, count(*) FROMhdr ; SELECT varno, avg(fg_depar), CORR(fg_depar, an_depar) FROM body WHERE fg_depar is NOT null ;

36. odbdiff • Enables comparison of two ODB databases for differences • A very useful tool when trying to identify errors/differences between operational and experimental 4DVAR runs • Usually a non-trivial task • Usage: odbdiff –q ‘SELECT …’ /dir1/DATABASE1 /dir2/DATABASE2 • By default the command brings up an xdiff-window with respect to differences • If latitude and longitude were also given in the data query, then it also produces a difference plot using odbviewer-tool

37. odbcompress • Enables to create very compact databases from the existing ones for • archiving purposes, or • for smaller database footprint (disk occupancy) • Makes post-processing considerably faster • The user can choose to • Truncate the data precision, and/or • Leave out columns that are less of an importance • Typical compression ratios vary between 2.5X … 11X • the high compression achieved for satellite data !!

38. odbdup/odbmerge • Allows f.ex. database sharing between multiple users • Over shared (e.g. NFS, Lustre, GPFS, GFS) disks • Duplicates [merges] database(s) by copying metadata (low in volume), but shares the actual (high volume) binary data • Also enables creation of time-series database, for example: odbdup –i “200701*/ECMA.conv” –o USERDB • The previous example creates a new database labelled as USERDB, which presumably spans over the all conventional observations during the January 2007 • The main point : user has now access to whole month of data as if it was a single database !!

39. odb2netcdf • Translates the result of a given ODB-query (or whole ODB-table) into a series of NetCDF-files, by default one file for each ODB data pool (i.e. partition) • Usage: odb2netcdf –q ‘SELECT …’ [-p pool_number] [-P] • The result files can be viewed with the standard NetCDF tools like ncdump and ncview • The files can also be created in the NetCDF packed format (caveat : truncated data precision), -P option was used

40. Some interesting facts on ODB • Written mainly in C-language • Except Fortran90-interface and IFS/4DVAR interface • Except BUFRODB (by Milan Dragosavac, ECMWF) • ODB/SQL is currently converted into C-code • 10 lines of SQL generates >> 100 lines of C-code • Standalone ODB installation (w/o IFS) is also available • Tested at least on the following machines • SGI/Altix, IBM Power3/4/5, Linux Intel/AMD • Fujitsu VPPs, NEC SX, Cray XT3/4 • Automatic binary data conversion guarantees database portability between different machines

41. … and some ODB “limitations” • ODB software is clearly meant for large scale computation since – given lots of memory and disk space, fast CPUs: • A single program can handle up to 2^31 ODB databases • A single database can have up to 2^31 data pools • A single database can have any number of tables • A single table in a data pool can have up to 2^31 rows and (by default) 9999 columns • A single ODB/SQL-query over active data pools can retrieve up to 2^31 rows in one go • These really big numbers show that ODBs potential is on parallel computers. Yet we haven’t forgotten the PCs!

42. Finally… • ODB software is developed to allow unprecedented amounts of satellite data through the IFS/4DVAR system • Software has been operational at ECMWF since June’2000, but is still evolving • Emphasis is now on graphical post-processing and how to enable fast access to very large amounts of data • Who is using ODB outside ECMWF ? At least … • MeteoFrance, Hungarian MS, SMHI, FMI • Aladin and some HIRLAM nations • Australian Bureau of Meteorology • University of Vienna via re-analysis ERA40 collaboration