Nandita Vijaykumar

  • Nandita Vijaykumar
    Nandita Vijaykumar

PhD Student, Carnegie Mellon University

Cross-Layer Compute and Memory Abstractions for Enhanced Programmability, Portability, and Performance in GPUs

Modern Graphics Processing Units (GPUs) have evolved into powerful programmable machines over the last decade, offering high performance and energy efficiency for many classes of applications. However, the performance that can be obtained from these systems depends on how the massively available parallelism and memory resources are managed by the programmer. There exists no powerful abstraction between the architecture and the programming model to easily manage these resources. This leads to underutilization of resources, and hence, a significant loss in performance when the application is not perfectly tuned for a given GPU. Even when an application is perfectly tuned for one GPU, our experiments show that there can still be a significant degradation in performance when running the same program on a different GPU. The existing interface also means that the management of resources is completely static, where there could be significant dynamic underutilization of resources at runtime, leaving significant performance on the table.

In this poster, I will present two ways to change this interface between the programming model and architecture in GPUs to enhance performance, ease of programming, and portability.

I am a PhD student in the Electrical and Computer Engineering Department at Carnegie Mellon University. I am advised by Prof. Onur Mutlu and Prof. Phil Gibbons.

My research interests lie in the general area of computer architecture and systems with a focus on the interaction between programming models, systems, and architectures, specifically in the context of memory systems and throughput-oriented accelerators such as modern GPUs. I am excited about codesigning programming models, systems and architectures with new richer interfaces between the layers of the computing stack to improve programmability and portability in achieving high performance.

During the course of my PhD, I’ve been fortunate to spend summers interning at Microsoft Research, Nvidia Research and Intel Labs.

Nandita Vijaykumar’s Research webpage
Research statement