Abstract: High-performance computing resources, such as multi-core Central Processing Units (CPUs) and Graphics Processing Units (GPUs) are readily available to programmers, due to the immense popularity of computationally demanding applications, such as machine learning, image processing, and scientific computing. Programmers today can access these resources in a variety of ways, including direct access and via the cloud. However, writing programs to utilize these resources efficiently and easily is a significant challenge.

 

In this talk, I will present new abstractions, domain-specific languages, and code generation techniques to efficiently program high-performance systems. I will present techniques to optimize image processing programs and scientific computing on modern GPUs that maximizes the data reuse opportunities by completely utilizing the deep memory hierarchy of modern GPUs. I will also present a domain specific language and a compiler that jointly optimizes communication and computation in distributed machine learning workloads. Finally, I will discuss new research challenges that must be solved to make efficient programming of high-performance systems easy and accessible to non-experts.

Short Bio: Abhinav Jangda is a final-year PhD Candidate at the University of Massachusetts Amherst advised by Prof. Arjun Guha. His research focuses on developing programming language abstractions and compilation techniques to help programmers leverage large scale systems efficiently. His work was invited for an article in USENIX :login;, has received an ACM SIGPLAN Distinguished Paper Award at OOPSLA, and a Best Paper Award at PACT.