Breadth-First Search (BFS)

Description: Measures performance for breadth-first search on an undirected (random, k-way) graph. Two algorithms are implemented. The first is from a team at Stanford, described in this paper: The second is from UIUC and detailed here: Some minor changes were made to these specifications for correctness and cross-platform compatibility.

Problem Sizes: (Graph vertices) - 1000, 10000, 100000, 1000000

Precision: N/A (graph data is unsigned int)

Includes PCIe Transfer Time: Yes, in [testName]_PCIe measurements


  • –graph_file instead of using a random graph, load one from a METIS file
  • –degree control the average degree of nodes, default 2
  • –algo select the algorithm to use, 1-IIT 2-UIUC
  • –dump-pl dump the path lengths to file
  • –source_vertex specify the vertex to start the search from, default 0
  • –global-barrier enable the use of a device global barrier in the UIUC algorithm. This can occasionally improve performance by reducing kernel launch overhead.

Specific Tests:

  • BFS - achieved memory bandwidth in GB/s
  • BFS_total - runtime in seconds (kernels and pcie)
  • BFS_kernel_time - runtime of all kernels in seconds
  • BFS_teps - achieved bandwidth in traversed edges per second, a common metric for graph algorithms
  • BFS_visited_vertices Reports the number of vertices visited in the traversal, should match the number of vertices in the connected component which contains the source vertex.
shoc/bfs.txt · Last modified: 2011/11/11 19:02 by kspafford
Recent changes RSS feed Driven by DokuWiki