DyCause

This repository is the source code of our paper "Faster, Deeper, Easier: Crowdsourcing Diagnosis of Microservice Kernel Failure from User Space" at ISSTA'2021.

ACM Reference Format:

Yicheng Pan, Meng Ma, Xinrui Jiang, and Ping Wang. 2021. Faster, Deeper, Easier: Crowdsourcing Diagnosis of Microservice Kernel Failure from User Space. In Proceedings of the 30th ACM SIGSOFT International Symposium on Software Testing and Analysis (ISSTA ’21), July 11–17, 2021, Virtual, Denmark. ACM, NewYork,NY, USA, 12 pages. https://doi.org/10.1145/3460319.3464805

Getting Started

Requirements

DyCause is written in Python 3.7.5. The dependencies are listed in environment.yaml. We recommand using anaconda to create an environment for testing DyCause. In current directory, run conda env create --file environment.yaml -n dycauseenv

Demo run DyCause

The entry code of DyCause is main_dycause_mp.py for multiprocess version or main_dycause.py for single process version. The multiprocess version is faster to test. The required parameters are documented in the code and can be showed by running with -h.

• For a demo execution of DyCause on PyMicro dataset, run:

  python main_dycause_mp.py pymicro 16 1 --start 1200 --step 30 --bef 100 --aft 0 --lag 9 --num_sel 1 --edge_thres 0.8 --verbose 2 --mean arithmetic

• For a demo execution of DyCause on our real-world microservice dataset, run:

  • (Multithread version) python main_dycause.py real_micro_service 14 6 28 30 31 --verbose 2
  • (Multiprocess version) python main_dycause_mp.py real_micro_service 14 6 28 30 31 --verbose 2

Demo run baselines

The baselines (TBAC, Netmedic, CloudRanger, MonitorRank, MicroCause) we implemented are also included here. Run python test_all.py to get a demo result for each baselines and DyCause.

For MicroCause, our implementation is in microcause.ipynb.

Detailed Description

We have tested each methods with different parameters to get reliable results.

PR@k and RankScore

For the methods TBAC, Netmedic and MonitorRank, we test them in parameter_tune.ipynb. For the methods CloudRanger and DyCause, we first generate the experiment results using cloudranger_params_tune.py and dycause_params_tune.py, and then analyze the results.

We provide the code to analyze the data and reproduce the figures (Figure 5, 7, 8, 9, 10, 11) in our paper in dycause_exp.ipynb. Also, we provide our experiment record files in the folder exp_records. Using these files, one can get the same preformance results as we do.

Efficiency tests

In order to test the efficiency of different methods, we also record the running time in some of our experiments. And we use the code in dycause_exp.ipynb to analyze them and generate the running time figures in our paper.