updates on parallel gcam · 2014-09-30 · 5 parallel structure of gcam firstlevel’supervisor’...

Post on 15-Aug-2020

9 Views

Category:

Documents

0 Downloads

Preview:

Click to see full reader

TRANSCRIPT

Joint GCAM Community Modeling Meeting and GTSP Technical Workshop Joint Global Change Research Institute College Park, Maryland, USA

Updates on Parallel GCAM

ROBERT LINK

1

!   Why do we need a parallel GCAM? !   How did we parallelize GCAM? !   What can you do with parallel GCAM? !   Some results from cost minimization calculations !   Future plans for parallel GCAM !   Discussion

2

Summary

3

Why Parallelize GCAM? (Why don’t you just run hundreds of copies?)

Policy  Run  

Quadrature  Sample  

Quadrature  Sample  

Quadrature  Sample  

Quadrature  Sample  

Quadrature  Sum   Policy  Cost  

Op7miza7on  Algorithm  

Parameter  Sample  

Parameter  Sample  

Parameter  Sample  

Parameter  Sample  

Op7miza7on  Algorithm  

Parameter  Sample  

Parameter  Sample  

Parameter  Sample  

Parameter  Sample  

Op7mal  Path  

Design of Parallel GCAM !   Builds on the parallel scenario evaluator

that was announced in the latest release !   Supervisory algorithms manage pools of

worker processes through a messaging protocol

!   You can write supervisors as necessary to do whatever kind of calculations you need to do.

Mul7-­‐threaded  GCAM  Worker  

Mul7-­‐threaded  GCAM  Worker  

Mul7-­‐threaded  GCAM  Worker  

Mul7-­‐threaded  GCAM  Worker  

Supervisor  

Messaging  protocol  used  to  pass:  •   Configura7on  •   Requests  for  return  data  •   Requests  to  evaluate  scenarios  •   Results  from  workers  

5

Parallel structure of GCAM

First-­‐level  Supervisor  Mul7-­‐threaded  GCAM  Worker  Mul7-­‐threaded  GCAM  Worker  Mul7-­‐threaded  GCAM  Worker  Mul7-­‐threaded  GCAM  Worker  Mul7-­‐threaded  GCAM  Worker  Mul7-­‐threaded  GCAM  Worker  

First-­‐level  Supervisor  Mul7-­‐threaded  GCAM  Worker  Mul7-­‐threaded  GCAM  Worker  Mul7-­‐threaded  GCAM  Worker  Mul7-­‐threaded  GCAM  Worker  Mul7-­‐threaded  GCAM  Worker  Mul7-­‐threaded  GCAM  Worker  

First-­‐level  Supervisor  Mul7-­‐threaded  GCAM  Worker  Mul7-­‐threaded  GCAM  Worker  Mul7-­‐threaded  GCAM  Worker  Mul7-­‐threaded  GCAM  Worker  Mul7-­‐threaded  GCAM  Worker  Mul7-­‐threaded  GCAM  Worker  

First-­‐level  Supervisor  Mul7-­‐threaded  GCAM  Worker  Mul7-­‐threaded  GCAM  Worker  Mul7-­‐threaded  GCAM  Worker  Mul7-­‐threaded  GCAM  Worker  Mul7-­‐threaded  GCAM  Worker  Mul7-­‐threaded  GCAM  Worker  

First-­‐level  Supervisor  Mul7-­‐threaded  GCAM  Worker  Mul7-­‐threaded  GCAM  Worker  Mul7-­‐threaded  GCAM  Worker  Mul7-­‐threaded  GCAM  Worker  Mul7-­‐threaded  GCAM  Worker  Mul7-­‐threaded  GCAM  Worker  

Second-­‐Level  Supervisor  

6

Example Configurations for two Parallel GCAM Calculations

Cost  calcula7on   Op7miza7on  

!   Objective: Find the minimum cost for a policy satisfying a constraint !   Constraint: 550 ppm at century end !   Configuration: 1920 processors for ~60 hours

7

Some Results from Cost Minimization Runs

!   Refining and fine-tuning supervisory algorithms !   Improve performance and parallel efficiency !   Constrain temperature, rather than concentration !   Include effects of uncertainty (new supervisory algorithm) !   Allow policies to respond to observed climate outcomes

8

The Road Ahead

!   This work was supported by the DOE Office of Science IARP

Acknowledgements

Discussion

10

top related