Multi-Population GA TSP

Description

This is a copy of SGA (TSP) which executes 10 SGAs side-by-side without migration. This workbench is an example for global parallelization with a distributed engine. However it is also possible to copy the whole thing into a sequential- or thread-parallel-engine.

Parameters

• Populations = 10
• Population Size = 100
• Mutation Rate = 5%
• Maximum Generations = 1000
• Elites = 1
• Parents = 200
• No Migration

Implementation & Notes

Works with revision r3.

Results & Conclusions

• Algorithmic results see: SGA (TSP).
• Effects of the parallelization: Speedup should be OK since we have only global parallelization and no communication between the populations. The DistributedEngine could be more efficient if we upload only he necessary parts of the scope tree to the executing client instead of the whole tree.

We tested the engine with following setup: DistributedEngine on user's computer. Grid-Server and client 1 on one machine of the grid. Clients 2 - 5 each on a different machine of the grid. Results of the test with the grid:

• Local (Avg. 96s 100%)
  1. 96s
  2. 94s
  3. 96s
• 1 Client (Avg. 86.6s 90% (caused by better performance of grid machine))
  1. 85s
  2. 89s
  3. 86s
• 2 Clients (Avg. 48s 50%)
  1. 49s
  2. 47s
  3. 48s
• 5 Clients (Avg. 22s 23%)
  1. 22s
  2. 23s
  3. 21s

---