Vapro

Vapro is a light-weight performance variance detection and diagnosis tool without requiring the source code of applications. It is practical for production-run parallel applications.

Get started

Dependencies

MPI consistent with applications
jsoncpp
papi
libunwind

Build from source

modify CPU_FREQ to real value in papi_wrap.cpp
compile as

mkdir build && cd build
cmake ..
make

Usage

We can use Vapro easily with two steps, profiling and analysis.

Profiling target applications

We can enable Vapro by setting LD_PRELOAD to preload the Vapro library before applications start.

export LD_PRELOAD=<path_to_libpapicnt.so>

Then, we can run the applications dierectly. Alternatively, we can make LD_PRELOAD only effective on the applications by

LD_PRELOAD=<path_to_libpapicnt.so> mpirun ./application_command

Analyzing results

Vapro saves results in the current working directory. There are four classes of files for different information.

log0_*: all calculation events
log1_*: all communication events
log2_*: relative performance data of calculation
log3_*: relative performance data of communication

Asterisks in the filenames are corresponding MPI ranks.

Roadmap

Supported CPU backends
- ✔ PAPI
- ✔ Linux perf
- ⬜ pmu-tools for Intel CPU
Supported GPU backends
- ⬜ CUDA
⬜ Integrated visualization

Contributor Guide

The development of Vapro is based on pull requests on Github. Before requesting for merging, a PR should satisfy the following requirements

Receive at least one approval from reviewers.
PR title should be concise since it is going to be the commit message in the main branch after merging and squashing.

Reference

Please cite our papers in your publications if they help your research:

@article{zhai2022detecting,
  title={Detecting Performance Variance for Parallel Applications Without Source Code},
  author={Zhai, Jidong and Zheng, Liyan and Zhang, Feng and Tang, Xiongchao and Wang, Haojie and Yu, Teng and Jin, Yuyang and Song, Shuaiwen Leon and Chen, Wenguang},
  journal={IEEE Transactions on Parallel and Distributed Systems},
  volume={33},
  number={12},
  pages={4239--4255},
  year={2022},
  publisher={IEEE}
}

@inproceedings{zheng2022vapro,
  title={Vapro: Performance variance detection and diagnosis for production-run parallel applications},
  author={Zheng, Liyan and Zhai, Jidong and Tang, Xiongchao and Wang, Haojie and Yu, Teng and Jin, Yuyang and Song, Shuaiwen Leon and Chen, Wenguang},
  booktitle={Proceedings of the 27th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming},
  pages={150--162},
  year={2022}
}

thu-pacman/VAPRO