Support for Debugging CUDA Apps Using ARM64 and NVIDIA GPUsBill Burns | December 3, 2019
Powerful CUDA Debugging Support
TotalView by Perforce now supports debugging CUDA applications utilizing NVIDIA GPUs on ARM64 servers. This added capability builds on TotalView’s already powerful CUDA debugging support for the Linux x86 64-bit and Linux PowerLE (Power8 and Power9) platforms.
ARM64 and GPU Debugging Support Now Available
Adding ARM64 and GPU debugging support was a fairly easy development effort, due to the consistent use of the CUDA debugging API across platforms. Leveraging TotalView’s existing debugging support for Linux ARM64 technologies, including Marvell’s ThunderX2, the TotalView development team was able to focus on enabling GPU debugging support for the ARM64 platform. Initial development was done on the ARM64/Volta GPU-based NVIDIA® Jetson AGX Xavier™ system.
Jetson AGX Xavier™ offers amazing compute capabilities at the fraction of the size and cost of a workstation, making it a perfect fit for a variety of autonomous machine applications. It’s also a great representation for the coming generation of ARM64/GPU compute servers. The development team was able to easily add CUDA GPU debugging support by simply enabling it within ARM64. All CUDA debugging API calls worked the same as other platforms and debugging of CUDA regions running on the GPU kernel worked as expected. The only challenge encountered during the porting was the requirement to debug targets as root or sudo; this is a limitation of the current Xavier solution and not an issue in ARM64/GPU server platforms. The TotalView solution was tested on ARM64/GPU compute servers and provided the same level of high performance GPU debugging support.
CUDA Developers Continue to Rely on TotalView
With the addition of GPU debugging support on ARM64, CUDA developers can continue relying on TotalView to provide the best GPU debugging solution for their applications. TotalView will continue aggressively moving forward in supporting CUDA advancements across all supported NVIDIA GPU-based architectures, and ensuring advanced debugging support for a variety of CUDA-based applications — from autonomous machine applications running on Jetson Xavier, to large scale multi-GPU HPC applications.
To learn more about the new TotalView 2019.3 features, read our recent press release.