Presentation is loading. Please wait.

Presentation is loading. Please wait.

Experiences and Decisions in Met Office coupled ESM Development

Similar presentations


Presentation on theme: "Experiences and Decisions in Met Office coupled ESM Development"— Presentation transcript:

1 Experiences and Decisions in Met Office coupled ESM Development
Richard Hill, February 2013 © Crown copyright Met Office

2 Contents Where did we come from? Where are we now?
Where do we want to go? Questions © Crown copyright Met Office

3 1: Where did we come from? 2004: A single exec with plenty of bottlenecks
Run the UM on n PEs Atmos Ocean Sea ice The Unified Model © Crown copyright Met Office

4 What Changed? Adoption of NEMO ocean and CICE sea-ice models
Not owned by Met Office! Code management trickier Not suitable for the single executable approach NEMO and CICE use of tri-polar grids! In-house coupling much harder Desire for more “plug compatible” components and higher resolution OASIS4 (initially) to the rescue! © Crown copyright Met Office

5 MPMD Coupling Code configuration influenced by:
NEMO and CICE MUCH FASTER than atmosphere NEC-SX6 vector machine NEMO and CICE essentially use the same tri-polar grid Not necessarily the same decomposition Coupling frequencies. NEMO  CICE: Every time step. NEMO-CICE  Atmosphere: Once a day Led to: NEMO-CICE combined in single executable © Crown copyright Met Office

6 OASIS3 Coupling “Upgrading” to OASIS3 (~2007)
Switch to stable (but serial) OASIS3 for production Pseudo-parallel OASIS3 Relatively straightforward to adapt components Tricky bits: Lack of effective debuggers When something’s wrong it’s not always obvious what! Creating, debugging and managing control files. “Framework” issues Control files (OASIS “namcouple” control, regridding weights etc) add to configuration and system management costs © Crown copyright Met Office

7 2: Where are we now? Current system.
Adapted UM control system in lieu of true coupling “framework”. (GUI, pre-processing, post-processing, data archiving.) Designed from the “inside-out” Pros: No huge (and potentially speculative) up-front development cost; Develop/extend as required; Intuitive for existing users and developers; All model definition info in a single place (traceability); building on existing well worn procedures for exchange and development of models. Used successfully by external partners. Cons: Modifying existing systems beyond original design constraints; Crowbar NEMO and CICE into system designed purely for the UM; Not inherently extensible; More components stretch infrastructure to/beyond limits. © Crown copyright Met Office

8 Current OASIS3 models Routinely run configurations (Typically 3-hour coupling) For Higher resolutions and coupling frequencies, e.g. N512-OASIS3-ORCA025 (The largest coupled model routinely run.) OASIS3 costs become more evident But still not crippling! Component memory increasingly a limitation on load balancing i.e. Size of model may inhibit ability to balance for optimal use of resources! © Crown copyright Met Office

9 Issues with direct bespoke coupling as used in NEMO-CICE
Bespoke direct coupling: May be quicker than going via separate coupler But ongoing code maintenance costs are significant! Extra responsibility of developing, maintaining, upgrading code which is owned by someone else. Little chance of getting local interfacing code into someone else’s code repository! Coordinating coupled model releases featuring local modifications, while keeping up to date with official component releases takes a lot of work. © Crown copyright Met Office

10 To attach or detach components?
E.g.: The way CICE/Internal coupling throttles NEMO © Crown copyright Met Office

11 Recent developments Installation of OASIS3-MCT
Upgrade costs lie in alterations to “framework” and management processes. No separate coupler exec Simplifies resource allocation, load balancing AND to some extent, management of control files (since no pseudo- parallel cases need to be considered.) Long parallel assessment climate runs already performed So far so good – coupling costs appear lower than OASIS3 Components still the limiting factor – not the coupler! 2nd Order Conservative remapping important? © Crown copyright Met Office

12 3: Where do we want to go? Increased resolution
N512-ORCA 1/12th global Will be a challenge (~4000+ cores probably.) Remapping weights files (SCRIP format ) at high resolutions are a worry – size, generation time and manipulation? OASIS3-MCT results so far suggest that the components continue to be the bottlenecks, not the coupler! Increased component numbers ESM plans to include wave, biogeochemical, ice sheet models, etc. Make the “right” decisions about generic/bespoke coupler use Load balance an N component system! © Crown copyright Met Office

13 Future models From this.....
Chemistry Atmos Coupler NEMO Land CICE © Crown copyright Met Office

14 Future shock To this.....? Wave IO Server Chemistry Atmos Coupler NEMO
Land CICE Biological © Crown copyright Met Office

15 Future N-component ESMS
Detached components may be more attractive: HPC architectures have changed since original decisions Optimal core count for each component more achievable??? More flexibility to load balance? Combined exec performance restricted by least scalable component e.g. component X may have scope for increased scalability while component Y doesn’t (or even reverses gains made by X.) Larger (higher res) models make NEMO-CICE a more significant cost (too many cores for CICE?) Could a more generic framework help development, usability, traceability, system and configuration management? © Crown copyright Met Office

16 Questions and answers © Crown copyright Met Office


Download ppt "Experiences and Decisions in Met Office coupled ESM Development"

Similar presentations


Ads by Google