DependableSystemsLab / GPU-Trident

GPU version of Trident. For more info see website. If you find this useful, please cite the paper.

GPU-Trident

GPU-Trident analytically models error propagation in CUDA kernels and predicts the SDC probability of these kernels and their instructions. It consists of LLVM compiler passes and python scripts.

Prerequisites

Dependencies (Tested on Ubuntu 14.04.6 LTS)

1. Python 2.7.6 or higher
2. NVCC v6.0.1
3. LLVM 3.0

Preparation

1. Add path to NVCC binaries to your system path and update the build path of LLVM (LLVM_PATH) in config.py.
2. Add the header #include "record_data.cu" in your benchmark file.
3. Annotate the kernels with bambooLogKernelBegin() before every call of your kernel, bambooLogRecordOff() after every kernel call of your kernel and bambooLogKernelEnd() after all invocation of that particular kernel. bambooLogKernelBegin() takes its dynamic call count as its input. Examples of annotation can be seen in example applications in ./benchmarks directory.

Execution

1. Populate the config.py and config_gen.py files according to kernel properties.
2. Run python prepare.py index command to get the readable_indexed.ll file. Use this file to get the index of output stores for this kernel and update the config.py file.
3. Run python prepare.py profile and populate domi_val in config.py, using ./results/lucky_store_details.txt file and results/loop_terminating_cmp_list file by using following format for each loop breaking cmp instruction in kernel index of branch breaking loop index of cmp breaking loop.
4. Run python prepare.py execute to get the overall SDC percentage and instruction SDC probability predictions in ./results/prediction.results file after command completion.

Benchmarks

Each directory in ./Benchmarks contains all the annotated source files and populated configuration files for example benchmarks. results_completed directory in each sub-directory of ./Benchmarks contains results obtained by running GPU-Trident on these applications.

To reproduce the results for these benchmarks perform the following steps:

1. Copy all the files and results directory from any sub-directory of ./Benchmarks to the root of GPU-Trident.
2. Run python prepare.py index; python prepare.py profile;python prepare.py execute command to get the SDC probability prediction results in ./results/prediction.results.