Séminaire KrakOS - Arkaprava Basu

Abstract

An ever-growing number of applications today rely on Graphics Processing Units (GPUs) for their computation. Consequently, the quality of GPU programs has an overarching impact on the reliability and efficiency of a large part of today’s software ecosystem. In this talk, we will discuss two aspects of GPU program quality – software reliability and efficiency.  We will show how writing correct and efficient GPU programs can be hard, even for the ninjas.

In the first part of the talk, we will discuss how subtle synchronization bugs in GPU programs can lead to intermittent failures, threatening the reliability of GPU-accelerated software.  We will then discuss our tool, named iGUARD,  that can pinpoint such synchronization bugs to programmers. In the second part of the talk, we will discuss how sub-optimal programming can lead to performance bugs in GPU programs, causing underutilization of costly and power-hungry GPUs. We will then discuss our second tool, ScopeAdvice, which can help programmers find performance bugs due to over and redundant synchronization in GPU programs. Together, these tools can help improve the quality of GPU programs.

Biography

Arkaprava (Arka) is an associate professor at the Indian Institute of Science (IISc), Bangalore, where he convenes the Computer Systems Lab. He is at EPFL, Switzerland, until July 2025 as a visiting professor (on sabbatical). Arka’s current research projects focus on the GPU’s compute stack – from the compiler, runtime, and device drivers to the GPU micro-architecture. Before moving back to India, Arka was researching with AMD Research in Austin, where he contributed to the US Dept of Energy’s Exascale projects that culminated with the Frontier supercomputer, the first Exascale supercomputer. Arka was inducted into the MICRO Hall of Fame in 2024. He received the ACM India Early Career Research Award (2025), Intel Rising Star Award (2021), and Google Faculty Awards in 2021 and 2023.