We provide the source code and the setup for iGUARD, a tool to detect races in GPU programs. iGUARD instruments GPU programs to detect races in them. It uses NVIDIA's NVBit [1], a GPU binary instrumenter, as the framework for instrumentation.
This README provides a peek into different parameters of the tool and a very high-level view of source code organization.
For full details refer to our paper: [Paper] [Video]
iGUARD: In-GPU Advanced Race Detection Aditya K Kamath and Arkaprava Basu ACM SIGOPS 28th Symposium on Operating Systems Principles (SOSP), 2021 DOI: https://doi.org/10.1145/3477132.3483545
To replicate the results given in the paper, we provide precompiled application binaries. Full details are given in the README in the benchmarks folder.
iGUARD is built on top of NVBit (version 1.4) and shares its requirements, listed below:
- SM compute capability: >= 3.5 && <= 8.0
- Host CPU: x86_64, ppc64le, aarch64
- OS: Linux
- GCC version : >= 5.3.0 for x86_64; >= 7.4.0 for ppc64le and aarch64
- CUDA version: >= 10.1
- CUDA driver version: <= 450.00
Currently no Embedded GPUs or ARMs host are supported.
Required software packages can be installed through apt using the following command:
sudo apt-get install -y wget bc gcc time gawk libtbb-dev
To download NVBit and compile the detector, run the following command:
make setup
The compiled detector can be found at nvbit_release/tools/detector/detector.so.
Once compiled, the detector can be run on binaries containing NVIDIA GPU code by setting the LD_PRELOAD environment variable. For example, to run the detector on an application binary called app.exe contained in the main repository folder, you would run the following command:
LD_PRELOAD=./nvbit_release/tools/detector/detector.so ./app.exe
Compiling the application binary with -lineinfo flag allows iGUARD to output line numbers when races are detected. Otherwise, SASS offsets are used.
We provide several knobs that allow users to change how the detection works. The major ones are listed below:
- BYTE_GRAN: The granularity of a single data item (in bytes) considered by the detector. (default = 4)
- CONT_OPT: Whether to perform contention optimizations (exponential backoff and detection coalescing) during race detection. (default = 1)
- EXIT: Whether to quit on encountering the first race. (default = 0)
- TIMEOUT: Time-out in seconds after which the application is terminated. 0 means never. (default = 0)
To use these knobs, set them as environment variables when performing race detection. For example, to time-out after 4 seconds when performing detection on app.exe, we run the following command:
TIMEOUT=4 LD_PRELOAD=./nvbit_release/tools/detetor/detector.so ./app.exe
The source code for the iGUARD race detector is found in the detector/ folder. The major files are as follows:
- detector.cu: This contains the CPU-side code for the detector. This includes allocating memory for metadata, the binary instrumentation process, and outputting caught races to the user.
- inject_funcs.cu: This contains the CUDA code run on the GPU after instrumentation. This includes incrementing relevant counters on synchronization operations, adding/removing locks from the lock table, and performing the in-GPU race detection.
[1] NVBit [Paper] [Repository]