Climate models and open source software Andy Pitman

Climate models and open source software Andy Pitman and Steven Phipps Climate Change Research Centre University of New South Wales

Background • The Earth’s climate is changing

Background 385 ppmv – increased from 280 in ~100 years 260 -280 ppmv over 10, 000 years

Temperature • Continued greenhouse gas emissions 2008 at 2007 or above current rates would 2006 cause further warming - very likely 2005 larger than those observed during the 20 th century.

How do we know ? • Observations – radiation changes – temperature changes – pressure, humidity, acidity etc • Basic theory • Backs of envelopes … and • Climate models

Climate models • Mathematical representations of – Atmosphere – Ocean – Terrestrial surface – Cryosphere – Fluxes of energy, mass and carbon between these systems

Climate Modelling Governing equations Forcing conditions Initial conditions Model output

Three-Dimensional Climate Models (GCM)

Climate models • ~300 x 300 km grid • ~20 vertical levels in atmosphere, 30 in ocean – 300, 000 grid points • Solves basic dynamics equations, parameterizes physics • Solve all equations with a ~15 minute timestep for 200 year integrations • Something like 2 x 1018 calculations per run

Climate models work Rainfall [annual] Which is observed and which is modelled ?

Climate models

Climate models • Written in Fortran – – Mainly f 95 or f 90, some f 77 best language [efficiency and safety] python, c, other wraps Mostly written by scientists not software engineers • Cannot [usefully] re-write in other languages • Run on NCI, or similarly large systems – main rationale for a Pflop system in Australia • Generate ~5 Tb per run. Data sets now > 1 Pbyte

Model Nationality BCC China BCCR Norway CCSM USA CGCM Canada CNRM France CSIRO Aus ECHAM Germany ECHO-G Germany FGOALS China GFDL USA GISS USA INM Russia MIROC Japan MRI Japan PCM USA UKMO UK Open source ?

Model Nationality BCC China BCCR Norway CCSM USA CGCM Canada CNRM France CSIRO Aus ECHAM Germany ECHO-G Germany FGOALS China GFDL USA GISS USA INM Russia MIROC Japan MRI Japan PCM USA UKMO UK Open source ?

Model Nationality BCC China BCCR Norway CCSM USA CGCM Canada CNRM France CSIRO Aus ECHAM Germany ECHO-G Germany FGOALS China GFDL USA GISS USA INM Russia MIROC Japan MRI Japan PCM USA UKMO UK Open source ? Via license, never latest version

Model Nationality BCC China BCCR Norway CCSM USA CGCM Canada CNRM France CSIRO Aus ECHAM Germany ECHO-G Germany FGOALS China GFDL USA GISS USA INM Russia MIROC Japan MRI Japan PCM USA UKMO UK Open source ? Via license, never latest version A variant may be available Via license, never latest version

Model Nationality Open source ? BCC China BCCR Norway CCSM USA CGCM Canada CNRM France CSIRO Aus ECHAM Germany ECHO-G Germany FGOALS China GFDL USA A variant may be available GISS USA Yes – fully accessible INM Russia MIROC Japan MRI Japan PCM USA UKMO UK Yes – fully accessible Via license, never latest version A variant may be available Via license, never latest version

Why ? • Not resourced or requested to provide this code in suitable standard • Costs of supporting open source – – – hardware compatibility complier variability, user support problems Code maintenance Dispute resolution/managing expectation – It’s a myth that climate modeling groups are well resourced … • Many climate models are variants of forecasting systems – commercial value

Why ? • Groups sometimes associated with defense departments • Culture • Fear of inappropriate use • IP concerns • Unwillingness to be exposed to independent evaluation • Strategic issues with best models • Why would you provide open source ?

So why open source? • Bugs per 1000 lines of code: – Commercial software: • ~1 -7 [Jones, 2000] • ~20 -30 [Cy. Lab, 2004] – Open source software: • 0. 17 [Linux kernel, Coverity, 2004] • 0. 434 [open source projects, Coverity, 2006] • So open source more reliable; better

So why open source? • A state of the art climate model has: • ~1 million lines of code • a software development lifecycle that lacks: – – external testing, comprehensive internal testing error checking. Analagous to commercial software? • contains ~1000 -30, 000 bugs?

So why open source? • National Institute of Standards & Technology, USA, 2002 – software errors cost US economy US$59. 5 billion/yr – Despite industry practice of spending 80% of development costs on identifying and correcting faults – Most modeling groups would not have one person dedicated to identifying and correcting faults • An open source climate model has a community dedicated to identifying and correcting faults

The climate model challenge • More components • Longer simulations • Finer spatial and temporal resolution • Serious computational challenges • 6 months to port to new computer system • Super computer time valuable

What I want … • model test plans with pre-defined performance – Software developers know how … • Code developed in a F 95 via subversion – You are welcome to try • Simple and elegant scripting – Suggestions welcomed • Tools to slice and dice net. CDF – Talk to the modelers – embed in R • Enhancements to open. DAP … • Implementation of agreed coding standards

Summary • Climate models have been developed over 20 years – with IT systems engineers in some cases • Few are open source – those that are were commonly developed with IT engineers • Hard to contribute – cannot ad hoc – little low hanging fruit • But a commitment in this area can be an invaluable contribution over time. Needs to be based on a dialog with the modelers