Examination of the WRF Model on Heterogeneous Distributed Systems

1

• Examination of the WRF Model on Heterogeneous
  Distributed Systems
• Wei Huang
• Jennifer Abernethy
• NCAR
• University of Colorado at Boulder

2
Outline

• Motivation
• Design and Implementation
• Experiments Results
• Conclusions
• Future Work

3
Motivation

• Weather models are computationally intensive
  applications
• Distributing the computation can increase
  resource utilization
• Increase computational power available from
  existing resources
• Run a larger domain
• Run at a higher resolution
• Enable the different architectures within each
  organization to work together
• Encouraging collaboration/cooperation between
  institutions

4
Distributed Heterogeneous Design

• Divide model domain into two parts
• Each sub-domain can run on different system with
  MPI
• Use TCP sockets to exchange data on overlapping
  borders of sub-domains

5
Distributed Decomposition
Domain
System 1
System 2
6
Distributed Heterogeneous Implementation

• One domain stands for server
• The other domain acts as client
• The server starts the socket, and listens to the
  client(s)
• The client connects to the server
• Information exchange between server and client
  every timestep
• 230 messages (1 per field)
• Each message is between 1-4KB

7
Distributed Communication

• Exchange information
• Client sends variables to server
• Server receives variables from client
• Server sends variables to client
• Client receives variables from server
• Server and client integrate forward 1 step
• Repeat the above procedure

(Figure to be inserted here)
8
Communication Pattern 1
9
Communication Pattern 2
10
Ways to Improve Performance

• Using derived data-type
• Using multiple threads to handle communication
• Increasing data size
• Pack message together
• Wider exchange boundary
• Dedicated node for remote message exchange
• Balanced CPU and I/O time on nodes

11
(No Transcript)
12
Distributing over Different Architectures

• PC Cluster 2 2.4GHz dual-node machines
• DEC alpha 4 667MHz processors
• (IBM 1.3 GHz power4)

13
Experiment Results
14
Conclusions

• WRF can run as a distributed heterogeneous
  application
• Main performance slowdown is amount of
  communication
• Bandwidth
• Amount of data

15
Future Work

• Dedicated Communication Node
• Balance between Communication and Computation
• Compare with Grid-enabled MPICH-G2 implementation

16
Questions?