Fractional 2-matching GPU solver
Fractional 2-matching (F2M) is a linear programming relaxation of the 2-matching problem which, in turn, is a relaxation of traveling salesman problem. Here we present a fractional 2-matching problem GPU solver based on gradient descent algorithm which exploits the special structure of F2M. The algorithm can be run in parallel mode and was implemented as CUDA C/C++ program.
Requirements:
You must (CUDA):
-
have a cuda device with Compute Capability at least 3.x.
-
have installed the CUDA Toolkit
-
have installed Thrust (the CUDA C++ template library)
You must (C++):
- have installed Boost
Compilation
To compile the code just run "make" command.
This would create a file: F2M.solver.exe.
Input/Output
To run program :
./F2M.solver.exe file N
You must provide 2 arguments:
- file is file in TSPLIB format
- N the number N (the top N closest neighbors for each node). Please note that for large scale problems (> 20,000 nodes ) the N should not exceed 200-500 or you risk to run out of memory (cuda device memory is commonly a limiting factor).
To run test example:
- ./F2M.solver.exe examples/vsli/xqf131.tsp 20"
As output, 2 files are provided:
- file with optimal solution: examples/vsli/xqf131.tsp.top20.F2M_solution
- file with optimal value : examples/vsli/xqf131.tsp.top20.F2M_report
The optimal solution for F2M is a set of edges with value 1 and a set of edges with value ½.
You can find a number of exapmles in the folders "examples/vsli/" and "examples/art/"