Automated software testing of memory performance in embedded GPUs

Abstract

Embedded and real-time software is often constrained by several temporal requirements. Therefore, it is important to design embedded software that meets the required performance goal. The inception of embedded graphics processing units (GPUs) brings fresh hope in developing high-performance embedded software which were previously not suitable for embedded platforms. Whereas GPUs use massive parallelism to obtain high throughput, the overall performance of an application running on embedded GPUs is often limited by memory performance. Therefore, a crucial problem lies in automatically detecting the inefficiency of such software developed for embedded GPUs. In this paper, we propose GUPT, a novel test generation framework that systematically explores and detects poor memory performance of applications running on embedded GPUs. In particular, we systematically combine static analysis with dynamic test generation to expose likely execution scenarios with poor memory performance. Each test case in our generated test suite reports a potential memory-performance issue, along with the detailed information to reproduce the same. We have implemented our test generation framework using GPGPU-Sim, a cycleaccurate simulator and the LLVM compiler infrastructure. We have evaluated our framework for several open-source programs. Our experiments suggest the efficacy of our framework by exposing numerous memory-performance issues in a reasonable time. We also show the usage of our framework in improving the performance of programs for embedded GPUs.

Performance and Security Analysis for GPU-Based Applications

“…On the contrary, we argue the importance of execution platforms in the context of detecting performance bugs, with a specific focus on GPUs. Previous works on performance testing (Chattopadhyay et al, 2014b;Banerjee et al, 2013) or other works on testing GPU programs (Li et al, 2012) concentrate on testinput generation and are not directly applicable for localizing the root cause of performance bugs. In this work, we aim to localize the cause of memory interferences by systematically comparing the original and the golden traces.…”

“…We emphasize that Therefore, we assume the presence of an existing test-suite, with which the GPU kernel can be executed. Such test suites can be generated via fuzzing the GPU kernel or via systematic testing (Chattopadhyay et al, 2014b). Broadly, we perform the following operations to localize the cause of memory interference across threads.…”

An Intelligent Waste Removal System for Smarter Communities

“…However, this has many algorithms and testing problems that need to be improved and optimized. As Zhen Peng points out, the performance and efficiency of embedded software architecture tends to decline in most applications, which delays the software development cycle [47].…”

A Search Based Approach for Stress-Testing Integrated Circuits

“…More work in [11] shows the use of SBST for worst case execution time analysis using multi-objective criteria. The possibility of automated test generation for testing non functional attributes of embedded systems is proven in [12], even though these don't seem to rely on search based techniques.…”