A Fast and Secure GPU Memory Allocator

Abstract

Graphics Processing Units (GPUs) is widely used to perform general purpose computing in many areas such as scientific computing and deep learning. In order to offer more flexibility in GPU programming, dynamic memory allocation has been introduced in GPU programming frameworks such as CUDA. However, the dynamic memory allocator in CUDA is inefficient in highly concurrent environments. Thus, several dynamic memory allocators are recently proposed to enhance the performance of dynamic memory management. However, these allocators only focus on achieving higher performance but ignore the security issues. In this paper, we propose a fast and secure GPU memory allocator based on ScatterAlloc. In order to efficiently protect against memory attacks such as buffer overflows, our allocator consists of several key techniques including canary-based memory protection (two options such as detection-on-free and always-on-detection are provided), address compression, and over-provisioning. Experimental results show that the allocator can effectively detect buffer overflow errors while it is still approximately 100 times faster than the CUDA toolkit allocator.