What is it about?
DrGPUM is the first memory profiler that systematically investigates memory inefficiency patterns in GPU-accelerated applications. DrGPUM utilizes NVIDIA Sanitizer APIs to gather crucial information on application memory behavior, such as GPU memory API invocation sequence and memory access locations. It finally categorizes the memory inefficiencies into either object-level or intra-object and provides rich memory optimization insights for developers.
Featured Image
Photo by Caspar Camille Rubin on Unsplash
Why is it important?
GPUs are extensively utilized in computing platforms to speed up applications across diverse domains. Nonetheless, the insufficient availability of GPU memory resources frequently hinders the ability to enhance the applicability of GPU computing. Hence, we proposed DrGPUM to profile and pinpoint the memory inefficiencies in GPU applications and offer developers insightful optimization suggestions. With the help of DrGPUM, we achieved significant peak memory consumption reduction in popular GPU benchmarks and applications.
Perspectives
DrGPUM supports object-level and intra-object analysis, which categorizes GPU memory inefficiencies into 10 inefficiency patterns. Moreover, DrGPUM targets fully optimized GPU binaries and does not require any modification of source code, which is compatible with production use. DrGPUM finally presents the profiling results in a user-friendly GUI that provides rich information about the inefficiencies and optimization suggestions of the monitored application. In a word, DrGPUM is a solid and useful GPU memory profiling tool for developers.
Mao Lin
University of California Merced
Read the Original
This page is a summary of: DrGPUM: Guiding Memory Optimization for GPU-Accelerated Applications, March 2023, ACM (Association for Computing Machinery),
DOI: 10.1145/3582016.3582044.
You can read the full text:
Contributors
The following have contributed to this page
