/Trident

Trident is a tool to analytically model the Silent-Data Corruption (SDC) rate of a program, to replace expensive fault injection experiments. Please refer to our papers at DSN 2018 for more details (links in the README)

Primary LanguageC++MIT LicenseMIT

Trident Project

Trident analytically predicts the SDC probability of a given program and its instructions without any fault injection. It is implemented as a set of LLVM compiler passes driven by python scripts. These passes are located in ./llvm-2.9/lib/Transforms/.

Getting Started

These instructions will get you a copy of the project up and running on your local machine.

Prerequisites

  1. The packages and dependencies required by LLFI(v2.9) and LLVM(v2.9)
  2. LLVM-GCC 4.2.1 (LLVM build 2.9)
  3. Python 2.7

Tested on Ubuntu 12.04 and 14.04 LTS

Preparation - Trident

  1. Configure the path for llvm_gcc binary (LLVMGCC) and Trident checkout (ROOT) based on your environment in ./test/Tridnet/config.py and ./inst_count/run_sampling.sh file.

  2. Put the compiled LLVM IR file of the program in the ./test/Trident and ./inst_count folder directories. The instructions should be indexed. For example, using instrument in LLFI. The file name should have following pattern <program_name>-profile_linked.ll

  3. Put the program input file (if any) in ./test/Trident and ./inst_count directories.

  4. Choose the LLVM IR instructions that are considered as the program output in ./test/Trident/getStoreMaskingRate.py (line:297).

  5. Update the paths, indexed LLVM IR file name (<program_name>), input command (<program_input>) and output file name (<output_file>) if any in ./inst-count/run_sampling.sh script according to the program being tested.

Execution - Trident

  1. Profile the dynamic footprint of the instructions of the program. This can be done in ./inst_count folder. Run the command bash run_sampling.sh. A log file called fi_breakdown.txt is generated.

  2. Copy the log file fi_breakdown.txt to the ./test folder. Trident will read and predict the SDC probabilities of the instructions provided in this log file.

  3. To execute Trident, run the command python prepare.py <program_name> "input command" in ./test/Trident folder. For example for blackscholes bench mark use the command python prepare.py blackscholes "1 in_4.txt output.txt".

  4. Trident will aggregate the overall SDC probability of the program based on their profiled dynamic footprint and store the result in ./test/Trident/prediction.results after command completion.

vTrident

vTrident is an extension of Trident. It simplifies the memory dependency sub-model of Trident to identify the variation of the SDC probablity of a given program with multiple program inputs. As a result, vTrident is much faster than either fault injection or Trident to bound the SDC probability of the program with multiple inputs.

Preparation - vTrident

  1. Profile the dynamic footprint of all the instruction of the program for all the inputs following steps provided for Trident. Rename the produced fi_breakdown.txt for each input as input_n-fi.txt where n is 0 and increments for each input.

  2. Place all the produced input_n-fi.txt files in ./test/vTrident/data directory.

  3. Put the compiled LLVM IR file of the program in the ./test/vTrident. The instructions should be indexed. For example, using instrument in LLFI. The file name should have following pattern <program_name>-llfi_index.ll

  4. If there are any program input files that are used in input command, place them in ./test/vTrident and if they have .txt extension, set the input_txt variable to 1 in ./test/vTrident/config.py .

  5. Choose the LLVM IR instructions that are considered as the program output in ./test/vTrident/getStoreMaskingRate.py (line:297).

  6. Configure the path for llvm_gcc binary (LLVMGCC),Trident checkout (ROOT), prgram_name used in step 3 (bmName) and input commands for all the inputs (inputList) in ./test/vTrident/config.py.

Execution - vTrident

To use vTrident execute the file ./test/vTrident/driver.py. After the execution of vTrident is complete, range of SDC and the input that can be used for finding median SDC using FI are provided in console and ./test/vTrident/results/variation.results.

Paper

Modeling Soft Error Propagation in Programs

http://blogs.ubc.ca/karthik/2018/03/04/modeling-soft-error-propagation-in-programs/

Modeling Input Dependent Error Propagation in Programs

http://blogs.ubc.ca/karthik/2018/03/04/modeling-input-dependent-error-propagation-in-programs/