GTrace

Code and dataset for (ESEC/FSE' 23) From Point-wise to Group-wise: A Fast and Accurate Microservice Trace Anomaly Detection Approach.

Evaluation of Accuracy

We provide dataset B for evaluation.

• Download dataset_b.zip and put it under dataset/ after unzipping it.
• Install Python 3.8+ on your system.
• Run pip3 install -r requirements.txt to install the dependencies.
• Run python3 -m tracegnn.models.gtrace.main to start training. The evaluation will automatically starts after training.
• If you want to run on GPU, you can modify the device in tracegnn/models/config.py.

(To enable accuracy evaluation, the model runs the inference once for each trace and thus cannot be used for time efficiency evaluation. If you need to perform evaluation of time efficiency, please refer to the following section.)

Evaluation of Time Efficiency

We provide the code for the Anomaly Detection module and Graph Building module. The code has not been tested on the Windows platform; please use a Mac or Linux system.

To evaluate the time efficiency, we provide a minimal example and a trained model that can be run directly on your local device without deployment:

• Run cd deployment.
• Install GCC 9.3.0+, make and CMake 3.2+ on your device. Run bash build.sh to download and build the dependencies.
• Run sh run_local.sh to evaluate the time efficiency.
• Install Intel SVML to get better performance on Intel CPU. (See https://numba.readthedocs.io/en/stable/user/performance-tips.html#intel-svml).
• When the the program is complete, it will enter a waiting state and will not exit automatically. Therefore, you need to manually kill it.

Visualization Tool

• Run python3 -m tracegnn.visualization.webviewer_server.
• Visit http://localhost:12312/0 or http://localhost:12312/1 to see the visualization results for two example cases.

Reference

• LRUCache11: https://github.com/mohaps/lrucache11.git
• Kubernetes: https://kubernetes.io/
• PyTorch: https://pytorch.org/
• DGL: https://dgl.ai/
• CMake: https://cmake.org/

Notes

• Dataset A is under review by the company and cannot be disclosed at this time.