This is the official implementation of "SPAtio-Temporal graph System (SPATS)" in the following paper:

• Yoon, Heeyong, Kang-Wook Chon, and Min-Soo Kim. "SPATS: a practical system for comparative analysis of spatio-temporal graph neural networks." Cluster Computing 28.13 (2025): 826, https://doi.org/10.1007/s10586-025-05523-6.

If you use this repository in your research project, please cite the following BiBTeX in your paper:

@article{yoon2025spats,
  author={Yoon, Heeyong and Chon, Kang-Wook and Kim, Min-Soo},
  title={SPATS: a practical system for comparative analysis of spatio-temporal graph neural networks},
  journal={Cluster Computing},
  year={2025},
  month={Sep},
  day={19},
  volume={28},
  number={13},
  pages={826},
  issn={1573-7543},
  doi={10.1007/s10586-025-05523-6},
}

⚠ This is not a commercial system, so we do not guarantee its behavior in all environments. Please handle exceptional cases with appropriate external knowledge, official documentation, and experience.

⚠ Some parts contain template text; do not copy and paste them directly, and make sure to replace the template text appropriately.

⚠ Because NFS recognizes users only by ID, it is strongly recommended to set all nodes' user and group IDs to the same value. Otherwise, unexpected behavior can occur (search about the Linux command id, usermod or groupmod).

1. Prepare your GPU cluster and set one node as master server and other nodes as worker server. The master server and worker server roles can be overlapped, but separating the roles is recommended because CPU usage might affect model training.

   Our testing environment for each node:

   • Ubuntu 18.04

   • CUDA 11.3 to 12.1 (different per node)

   • Python 3.10 (installed by miniconda) and necessary packages (PyTorch, Numpy, ...)

2. Install NFS server on the master server

   a. Make the repository visible to anyone.

      sudo apt install nfs-kernel-server -y # install NFS server

   b. Modify /etc/exports.

      sudo vim /etc/exports

   c. Write the following text line at the bottom of the /etc/exports/.

   /your_master_server_folder/SPATS/ *(rw,sync,no_subtree_check)

   d. Restart NFS.

      sudo exportfs -a
      sudo systemctl restart nfs-kernel-server

   e. Download this repository on the master server.

      cd /your_master_server_folder/
      git clone https://github.com/sunrise2575/SPATS

   f. Make the repository visible to anyone.

      sudo chmod 777 ./SPATS/ # Make the repository visible to anyone

3. Install NFS client and mount the remote folder on each worker server

   a. Install NFS client.

      sudo apt install nfs-common -y

   b. Mount the master server's folder.

      mkdir -p /your_worker_server_folder/SPATS/
      mount <master_server_IP>:/your_master_server_folder/SPATS /your_worker_server_folder/SPATS/

   For checking IP, use ip a or ifconfig -a command.

   c. Make sure the remote folder is mounted well.

      ls -al /your_worker_server_folder/SPATS/

4. Install Python dependencies on every server (all master server and worker server)

      conda activate base # or specify different virtualenv
      pip install -r requirements.txt

   Some Python packages in requirements.txt are not version-sensitive, so you can selectively modify requirements.txt not to upgrade or downgrade your existing packages.

1. Launch Broker and Worker

   a. Launch Broker process on master server

      python ./broker.py --port <broker_port>

   The default port number is 9999; if you do not specify the port number, you can type like this:

      python ./broker.py

   b. Launch Worker process on each worker server

      python ./worker.py --broker <broker_IP>:<broker_port> --gpu <GPU_indices_as_you_wish>
      # python ./worker.py --broker 1.2.3.4:9999 --gpu 0,2,5 # example; using 0,2,5-th GPU only

   We recommend using tmux to manage processes on each node efficiently.

2. Insert jobs to Broker

   ⚠ The following content may be lengthy and complex. Since this was developed for research purposes and aims to execute queries in Python without using a structured language like SQL, the details can be intricate. It is recommended to read through carefully and learn by using query-bulk-insert.py yourself.

   a. Find the following text lines in query.py and query-bulk-insert.py

      BROKER_IP='127.0.0.1'
      BROKER_PORT='9999'

   and replace it with the IP and port number of your Broker.

      BROKER_IP=<broker_IP>
      BROKER_PORT=<broker_port>

   b. Insert multiple jobs using query-bulk-insert.py

   In query-bulk-insert.py, the variable DEFAULT is a default job setting. By changing some values of DEFAULT and VARIATION, query-bulk-insert.py generates thousands of combinations of jobs.

   The following example shows how to use VARIATION properly.

      VARIATION = {
          'maxEpoch': [10],
          'batchSize': [64, 128], # 2 variations
          'datasetName': ['METR-LA', 'PEMS-BAY'], # 2 variations
          'modelName': ['Seq2Seq_RNN', 'ASTGCN'], # 2 variations
          'adjacencyMatrixThresholdValue': [float('0.' + str(i)) for i in range(2)], # 2 variations
          # -> 2x2x2x2=16 variations
      }

   In this case, SPATS sets maxEpoch to 10, while batchSize to 64 and 128 both. The candidate datasetName and modelName work similarly. You can write Python generator syntax on VARIATION such as adjacencyMatrixThresholdValue of the example.

   On the other hand, DATASET_DEPENDENT_SETTING and MODEL_DEPENDENT_SETTING in query-bulk-insert.py are variables to help change other values in DEFAULT comfortably, which are affected by datasetName and modelName in VARIATION, respectively. For example, the above VARIATION example specifies that candidate datasets are METR-LA and PEMS-BAY, following items in DATASET_DEPENDENT_SETTING are selected,

      DATASET_DEPENDENT_SETTING={
        ...
          'METR-LA': {
              'additionalTemporalEmbed': ['timestamp_in_day'], # 1 variation
              'targetSensorAttributes': [['speed']], # 1 variation
              'inputLength': [12], 'outputLength': [12], # 1 variation each
              # -> 1x1x1x1=1 variation
        },
          'PEMS-BAY': {
              'additionalTemporalEmbed': ['timestamp_in_day'], # 1 variation
              'targetSensorAttributes': [['speed']], # 1 variation
              'inputLength': [12], 'outputLength': [12], # 1 variation each
              # -> 1x1x1x1=1 variation
        },
        ...
       }

   Similarly, the following items in MODEL_DEPENDENT_SETTING are selected because the modelName is ['Seq2Seq_RNN', 'ASTGCN'].

      MODEL_DEPENDENT_SETTING={
        ...
          'Seq2Seq_RNN': { 'adjacencyMatrixLaplacianMatrix': [None] }, # 1 variation
          'ASTGCN': { 'adjacencyMatrixLaplacianMatrix': ['cheb_poly'], }, # 1 variation
        ...
      }

   If the default value is like this:

      DEFAULT={
          'trainTestRatio': 0.7,
          'adjacencyMatrixThresholdValue': 0.7,
   
          'maxEpoch': 100,
          'batchSize': 64,
   
          'lossFunction': ["MAE", "MSE", "MAAPE"], # this should be list; 1 variation
          'targetLossFunction': "MSE",
   
          # about optimizer
          'optimizer': 'Adam',
          'learningRate': 0.001,
          'weightDecay': 0.0005,
      }

   By the power of algorithm stdin_generator() and additional model configuration, which is stored in .yaml files, 16 variations of jobs are generated, such as

      stdin = {
          'datasetName': 'METR-LA', # set by VARIATION
          'adjacencyMatrixThresholdValue': 0.0, # changed by VARIATION
          'modelName': 'Seq2Seq_RNN', # set by VARIATION
          'maxEpoch': 10, # changed by VARIATION
          'batchSize': 64, # changed by VARIATION
   
          'adjacencyMatrixLaplacianMatrix': None, # set by MODEL_DEPENDENT_SETTING
          'modelConfig': {}, # automatically filled by common/model/<MODEL_NAME>.yaml. if it doesn't exist, it becomes an empty dict.
   
          'additionalTemporalEmbeds': ['timestamp_in_day'], # set by DATASET_DEPENDENT_SETTING
          'inputLength': 12, # set by DATASET_DEPENDENT_SETTING
          'outputLength': 12, # set by DATASET_DEPENDENT_SETTING
          'targetSensorAttributes': ['speed'],  # set by DATASET_DEPENDENT_SETTING
   
          'trainTestRatio': 0.7, # from DEFAULT
          'learningRate': 0.001, # from DEFAULT
          'weightDecay': 0.0005, # from DEFAULT
      }

   or

      stdin = {
          'datasetName': 'PEMS-BAY', # set by VARIATION
          'adjacencyMatrixThresholdValue': 0.1, # changed by VARIATION
          'modelName': 'ASTGCN', # set by VARIATION
          'maxEpoch': 10, # changed by VARIATION
          'batchSize': 128, # changed by VARIATION
   
          'adjacencyMatrixLaplacianMatrix': 'cheb_poly', # set by MODEL_DEPENDENT_SETTING
          'modelConfig': {
              'time_strides': 1,
              'nb_block': 2
              'nb_chev_filter': 64,
              'nb_time_filter': 64
        }, # automatically filled by common/model/<MODEL_NAME>.yaml. if it doesn't exist, it becomes an empty dict.
   
          'additionalTemporalEmbeds': ['timestamp_in_day'], # set by DATASET_DEPENDENT_SETTING
          'inputLength': 12, # set by DATASET_DEPENDENT_SETTING
          'outputLength': 12, # set by DATASET_DEPENDENT_SETTING
          'targetSensorAttributes': ['speed'],  # set by DATASET_DEPENDENT_SETTING
   
          'maxEpoch': 10, # changed by VARIATION
          'batchSize': 128, # changed by VARIATION
   
          'trainTestRatio': 0.7, # from DEFAULT
          'learningRate': 0.001, # from DEFAULT
          'weightDecay': 0.0005, # from DEFAULT
      }

   and so on. you can add print(stdin) at the loop of main() in query-bulk-insert.py to print generated stdin for better understanding and debugging

   c. After DEFAULT and VARIATION are ready, you can insert jobs and wait for the completion.

      python ./query-bulk-insert.py # run SPATS!

   Broker works as a queue, so you can insert more jobs without waiting for the completion of previously inserted jobs.

3. Get the job information and delete job

   a. Full job info. (only shows recently inserted 100 jobs)

      python ./query.py list

   b. Job info with filtered type (only shows recently inserted 100 jobs).

      python ./query.py list <started|pending|success|failure>

   c. Single job info.

      python ./query.py select <job_id>

   d. Delete a job.

      python ./query.py delete <job_id>

4. Visualize results

   a. Copy broker.sqlite3 related files to visualize/ folder

   cp broker.sqlite3* visualize/.

   b. Run extractor

   cd visualize/
   python ./extractor.py broker.sqlite3 success

   c. Run 1-replace-model-and-dataset.ipynb and 2-concat-results.ipynb in order to get result.csv

   d. Use make-fig*.ipynb files to generate comparison results similar to our paper.