AutoCheck

AutoCheck is a dependency analysis tool, which aims to autmatically identify objects for checkpointing for large-scale machine learning models and HPC applications.

Docker setup

We highly recommend that the reviewers use the docker container for NO prerequisite software installation, which could take a few hours to setup.

1. Docker Pull Command

docker pull zhangwping/vpautocheck:latest

2. Create container

docker run -it zhangwping/vpautocheck:latest /bin/bash

Install with source code

If you cannot use Docker, then read on. It may take a few hours to install the following prerequisite software.

Requirements:

1. LLVM-Tracer 1.2 (available for download here, and follow the instructions on the page to install).
2. CMake 3.5.1 or newer.
3. graphviz 2.38.0 or newer (available for download here).

Build:

1. Clone AutoCheck

  git clone https://github.com/zRollman/Autocheck.git
  cd AutoCheck/

2. Install

  mkdir build/
  cd build
  cmake ..
  make

Run:

We provide two scripts for testing (auto.sh and run.sh). With the auto.sh script all experiments performed in our paper are automatically executed and the results are stored in the result directory (/workspace/AutoCheck/build/result). With the run.sh script, you can follow the script prompts to enter the trace file location and the information about the location of the main computation loop to perform a single benchmark test, and the results will be stored in the result directory.

Test with auto.sh:

1. Go to Autocheck

cd /path/to/your/AutoCheck/build

2. Execute auto.sh

./auto.sh

Test with run.sh:

Example benchmark :Himeno

1. Go to Autocheck

cd /path/to/your/AutoCheck/build

2. Execute run.sh

./run.sh

3. Enter the name of the trace file to be tested

Himeno

4. Enable or disable OpenMP acceleration (on or off)

off

5. Enter the name of the function where the main computation loop is located

jacobi

6. Enter the location where the function for which the main calculation loop exists is called

116

7. Enter the line number at the beginning of the main calculation loop

186

8. Enter the line number of the end of the main calculation loop

217

Benchmark's main loop location

Benchmark	function	call location	start line	end line
Himeno	jacobi	116	186	217
HPCCG	HPCCG	138	118	146
CG	main	0	296	330
MG	main	0	266	276
FT	appft	74	109	119
SP	main	0	184	190
EP	main	0	168	213
IS	main	0	732	736
BT	main	0	180	186
LU	ssor	205	115	267
CoMD	main	0	180	186
miniAMR	ssor	205	115	267
AMG	main	0	462	553
HACC	main	0	318	523

Validation:

We provide a homemade library with C/R functionality, and we add C/R code to all 10 benchmarks. When executing our modified program, only need to add one parameter (0 to start the program normally, 1 to start the program from the checkpoint file) to invoke the C/R functionality.

To verify the correctness of the variables provided by AutoCheck for checkpointing, we run the same benchmark three times (twice with the C/R code added and once with the original benchmark). First, we execute the program with the C/R code which will be interrupted by the fail-stop failure we simulated, then we restart the program with the data of the critical variables we stored in memory. Finally, we execute the original program and verify whether the program restarted successfully by comparing the execution result of the original program with the result of the restarted program.

Example benchmark :Himeno

1. Go to Himeno's (with C/R enable) working directory and compile the executable file

cd /path/to/your/Benchmarks/checkpoint/Himeno/build

cmake .. && make

2. Execute Himeno

./Himeno 0

3. Restart Himeno

./Himeno 1

4. Go to the working directory of the original Himeno program and compile it to get the executable file

cd /path/to/your/Benchmarks/varify/Himeno && make

5. Execute Himeno

./Himeno