Tahoe: Tree Structure-Aware High Performance Inference Engine for Decision Tree Ensemble on GPU

Tahoe enables high-performance GPU inference for decision tree ensemble which can be used in high performance computing, advertising systems, medical diagnosis, financial fraud detection, etc.

The easiest way to test it

Run the SVHN data set as an example.

sh ./run_an_example.sh

Run the 15 data set. (a little bit time-consuming)

sh ./run_all_15_examples.sh

---------------MORE DETAILS FOR TESTING-----------------

Requirement (pass-test)

• Ubuntu 16.04
• CUDA 10.1

Download trained forest and inference data

# Take the SVHN data set as an example.
# Download trained decision tree model
https://drive.google.com/file/d/1Z04IvVIIZdjUGZGIf4CpUQ2Ll06JVVll/view?usp=sharing
# Download data set
https://drive.google.com/file/d/1sqnuc8O9c58qphMHYIaNZSiHdDKV0AMF/view?usp=sharing

# In particular, all 15 trained models and data can be downloaded here (not necessary for testing)
https://drive.google.com/drive/folders/1S9ohva-P6NPB2GW8E1kkMqpQ-KmLL-E0?usp=sharing

# If you want to train the model from scratch (not necessary for testing)
Please refer to our another repo: https://github.com/zhen-xie/Decision-tree-ensemble.git

Build/Install from Source

git clone https://github.com/zhen-xie/Tahoe.git
cd Tahoe;
make Tahoe;

Run an example

# Run Tahoe with the corresponding paths. The first parameter is the model path, and the second parameter is the data path.
./Tahoe [PATH_MODEL] [PATH_DATASET]

Output

[zxie10@gpu050 Tahoe]$ ./Tahoe /home/zxie10/Project1/model_SUSY.txt /home/zxie10/Project1/data_SUSY.txt;
Performance model choose #1 strategy.
/home/zxie10/Project1/model_SUSY.txt
Load forest
Load data
Predict on CPU and get standard results...
Test on GPU...
time_dense is 0.990169 us
results are correct
Using strategy 0
time_dense is 1.082617 us
results are correct
Using strategy 0
time_dense is 3.838786 us
results are correct
Using strategy 1
time_dense is 0.120020 us
results are correct
Using strategy 3
time_dense is 0.212903 us
results are correct
Enumerate al strategies and choose #1 strategy.
Performance model predicts correctly
Tahoe brings 8.25x speedup.

Tahoe also provides a C++ interface to load forest and data from files and computes the prediction on GPU:

# Setup input forest, dataset, and algorithm, and construct the interface of Tahoe framework
BaseTahoeTest* pTest = new BaseTahoeTest(argv[1], argv[2]);
# Launch the inference of tree traversal
pTest->SetUp();
# Free resources
pTest->Free();