The autotuner for GraphIt aims to automatically find the best schedules for a given input algorithm specification and input graph. The autotuner is built on top of OpenTuner to efficiently search for the best combination of schedules (optimizations). Currently, the autotuner still requires the user to place a label "s1" on the operator that nees to be tuned.
The autotuner is still very experimental. Please read the instructions below carefully before using it.
Please first install OpenTuner. GraphIt's autotuner is built on top of OpenTuner.
Currently it is the easiet to use the graphit files under autotune/apps because we need to have hardcoded starting points for the BFS and SSSP. If you want to change the starting point for BFS or SSSP, simply edit bfs_benmark.gt or sssp_benchmark.gt.
To tune the performance of bfs on the small test graph 4.el, use the following command. --algo_file specifies the algorithm file, and --graph specifies the input graph to tune on. The autotuner have options that enable/disable parallelization, enable/disable NUMA optimizations (by default we disabled NUMA optimization), and other options such as setting an upper bound on the number of segments to use for configNumSSG. We usually put a time limt on tuning with the --stop-after option. Even for programs that run on a large input graph, we can usually finish within 5000 seconds.
#change into the directory for autotuner
cd graphit/auotune
#autotune serial pagerank with 4.el graph for 10 seconds
python graphit_autotuner.py --enable_parallel_tuning 0 --algo_file apps/pagerank_benchmark.gt --graph ../test/graphs/4.el --stop-after 10
#autotune parallel pagerank with 4.el graph for 10 seconds
python graphit_autotuner.py --enable_parallel_tuning 1 --algo_file apps/pagerank_benchmark.gt --graph ../test/graphs/4.el --stop-after 10
To see all the options
python graphit_autotuner.py -h
The final result will be displayed as a configuration in the standard output, and also in final_config.json. One example is shown below. You can then translate that into the scheduling commands. At this time, this translation step still has to be done manually.
('Final Configuration:', {'parallelization': 'serial', 'direction': 'SparsePush', 'numSSG': 5, 'DenseVertexSet': 'boolean-array'})
You can also set the serial and parallel compiler used for the C++ files in the graphit_autotune.py file by modifying the variables serial_compiler and par_compiler.
Currently, the autotuner can not tune the data layout. So the user would need to provide schedules for fusing together different vectors. The user can provide some scheduling commands in a file as a necessary schedule using the --default_schedule_file option.