cudaFlow benchmarks

File Descriptions

• main/

  • sample_graph.cu: main source file to run different graphs using cudaFlow

• bencmarks/sample_graph/sample_graph/

  • extreme_graph/: graphs with sequential and embarassing parallelism
  • map_reduce/: graph with map_reduce (i.e. Diamond)
  • random_DAG/: graph with random_DAG
  • tree/: graph with tree
  • utility/graph_executor.hpp: An executor to map a graph into taskflow. Each node contains four tasks: H2D, reduce sum, D2H, and set visited.

Prerequisites

Nvidia CUDA Toolkit and Compiler (nvcc) at least v11.0 with -std=c++17
GNU C++ Compiler at least v8.0 with -std=c++17
CMake at least v3.9

Step 1: Compile

Out of soucr build:

~$ mkdir build
~$ cd build
~$ cmake ../
~$ make

Step 2: Run

~$ cd bin
~$ ./simple_graph -g Graph --a Args

- GraphType: SerialGraph, ParallelGraph, Tree, MapReduce, RandomDAG
- Args: 
  - SerialGraph (Sequential): number of nodes
  - ParallelGraph (Embarrassing Parallelism): number of nodes
  - Tree: degree per node, number of levels
  - MapReduce: number of partitions, number of iterations
  - RandomDAG: number of levels, maximum number of nodes per level, maximum number of edges per node
 
For example, ./simple_graph -g RandomDAG -a 10 5 8 
will launch a randomly produced DAG that has 10 levels. Each level has at most 5 nodes and each node has at most 8 edges.