VaniDL

VaniDL is an tool for analyzing I/O patterns and behavior with Deep Learning Applications. It analyzes Darshan Extended traces to extract various I/O patterns with Deep Learning applications. The tool at it core uses vanidl which converts Darshan Profiler’s trace into knowledge for analysis. It is designed to provide low-level I/O behavior details to tensorflow applications in HPC.

VaniDL features include:

Easy-to-use and understand high-level API for extracting I/O behavior of the applications.
Fast prototyping through highly modular data representation through pandas for easy plotting of graphs.
Full transparency over profiling data with access to internal data structures such as timeline of applications, aggregation functions, and drill up/down data views.
Powerful helper functions to build a visual understanding of how applications perform I/O such as request distributions, file access pattern, and extracting file specific summaries.
Easy to use File Summary and Job Summary extractors for understanding the data consumed by Deep Learning Applications

Overview

#Initialize class
import vanidl
from vanidl.analyzer import *
profile = VaniDL()
#Load darshan file
status = profile.Load("./run1.darshan")
#Get Job Summary
summary = profile.GetSummary()

#Application Timeline of data operations
tl = profile.CreateIOTimeline()
plt.figure(figsize=(20,4))
plt.grid()
plt.plot(tl['time_step'], tl['operation_count']);

timeline

More examples are here

Installation

Requirements

numpy==1.18.5
pandas==1.0.4
h5py==2.10.0
tensorflow~=2.2.0

VaniDL Installation

To install VaniDL, the easiest way is to run

For the bleeding edge version (recommended):

pip install git+https://github.com/hariharan-devarajan/vanidl.git

For the latest stable version:

pip install vanidl

Otherwise, you can also install from source by running (from source folder):

python setup.py install

On Theta

module load VaniDL

Getting Started

See Getting Started with VaniDL to learn about VaniDL basic functionalities or start browsing TFLearn APIs.

Examples

There are many examples of analysis available, see Examples.

Contributions

This is the first release of VaniDL, if you find any bug, please report it in the GitHub issues section.

Improvements and requests for new features are more than welcome! Do not hesitate to twist and tweak VaniDL, and send pull-requests.

For more info: Contribute to VaniDL.

License

MIT License