Welcome to Seastar’s documentation!
This documentation was made to understand how the Seastar compiler works. Feel free to use it for your project and learning purposes.
What is Seastar?
Seastar is a system for programming GNN models using a vertex-centric approach. This new programming approach ensures that the logic for various GNN models can be written intuitively. It also allows those reading the code to ascertain the models purpose. The Seastar system can be used to implement models like GCN and GAT with relative ease.
While Seastar provides a new programming approach, it can perform optimizations on the model to bridge some of the gaps in other GNN frameworks like DGL and PyG. Seastar takes a function representing the operations performed on a single vertex as input.
Seastar is successful in achieving lesser memory consumption and faster execution in comparison to PyG and DGL.
exp
exp (experiments) directory contains code for different GNN models like gcn, gat, rgcn and appnp with
which the datasets are trained on. Each of the model stated above is implemented using dgl, pyg and seastar.
The result directory has a csv file final_result.csv which contains the results of the experiment/training
conducted by each GNN model on the datasets.
APPNP
GAT
GCN
RGCN
run_exp
This is where the entire program starts from, when you run it on Google Colab or any other platform.
Note
This is not where Seastar starts from
create_exp_list_sample0()
- create_exp_list_sample0(args)
According to the different GNN models passed as input through the command line,
create_exp_list_sample0returns a list (exp_list) of GNNExp class objects.- Parameters
args – Contains the arguments passed through the command line. Arguments are models, systems, gpu and num_epochs. eg:
args = Namespace(gpu=0, models=['gcn'], num_epochs=200, systems=['dgl', 'seastar'])- Return type
List of GNNExp class objects. These objects can be from the following classes: GATExp, GCNExp, APPNPExp and RGCNExp
main()
- main(args)
Format the exception with a traceback.
- Parameters
args – Contains the arguments passed through the command line. Arguments are models, systems, gpu and num_epochs. eg:
args = Namespace(gpu=0, models=['gcn'], num_epochs=200, systems=['dgl', 'seastar'])- Return type
None
if __name__ == __main__:
parser = argparse.ArgumentParser(description='Experiment')
Within this if-block, the arguments which are passed through the command line are parsed and stored inside an argparse.ArgumentParser object.
parser.add_argument("--models", nargs='+', default='gat', help="which models to run. Example usage: --models gat gcn")
parser.add_argument("--systems", nargs='+', default='dgl', help="which models to run. Example usage: --systems dgl seastar pyg")
parser.add_argument("--gpu", type=int, default=0, help="which GPU to use. Set -1 to use CPU.")
parser.add_argument("--num_epochs", type=int, default=200, help="number of training epochs")
–models: Refer to the GNN models which are to be used for training eg: GCN, GAT, etc.
–systems: Refer to the frameworks which are used for implementing these models eg: DGL, PyG and Seastar
–gpu: To whether use a gpu or not
–num_epochs: The number of epochs required for training
args = parser.parse_args()
args.models = [args.models] if not isinstance(args.models, list) else args.models
args.systems = [args.systems] if not isinstance(args.systems, list) else args.systems
Parses the input command and stores the arguments in args . Also makes sure that args.models and args.systems are list datatype.
This is what args would look like for the following command line input:
>> ./run_exp.py --models gcn --systems dgl seastar
args = Namespace(gpu=0, models=['gcn'], num_epochs=200, systems=['dgl', 'seastar'])
Attention
Need to figure out nb_access and dataset directory
python
seastar
Compiler Interface (zoom_v2)
- class Context
GraphInfo is a named tuple data structure whose fields can be accessed using indexing and the following field names
GraphInfo(number_of_nodes, number_of_edges, in_row_offsets, in_col_indices, in_eids, out_row_offsets, out_col_indices, out_eids, nbits)- _f
- _nspace
- _entry_count: int
- _run_cb
- _input_cache: dict
- _graph_info_cache
- _executor_cache