Congratulations on your graduation Doctor Ananya Muddukrishna

Where are you from and where did you study before coming to KTH-ICT?

I was born in Mysore, India. I studied at a collegiate state university called Visvesvaraya Technological University in my hometown for a Bachelor of Engineering degree with specialization in Electronics and Communication Engineering. I came to KTH to study for a Master of Science degree specializing in System-on-Chip Design and went on to study for the PhD degree.
 

What is your topic and why did you choose it?

My topic is data locality optimizations and high-resolution performance analysis techniques for programs written in a parallel programming language called OpenMP.
I chose the topic because it enables me to explore and understand parallel computing and system design deeply.
An additional reason is that I believe in OpenMP's vision to be an easy-to-use, portable parallel programming model and wanted to solve productivity problems that nag OpenMP users.
 

Describe your topic in short for a person that doesn't know much about it.

Multicore processors are everywhere today. Programming multicore processors efficiently using a popular language called OpenMP is painful for programmers since tools do not connect performance problems to program structure and there is limited support for data distribution and locality-aware scheduling.

 

Tell me something about your results.

My dissertation contributes with a workflow that simplifies performance analysis and optimization of OpenMP programs for programmers. The workflow is based on a visualization method called the Grain Graph that pinpoints performance problems. Results show that grain graphs can reduce the semantic gap between program structure and performance problems in existing visualizations, and guide programmers to make portable optimizations without tedious trial-and-error tuning.

The dissertation also contributes with data locality optimizations that programmers can make while iterating the grain graph based workflow.
The optimizations are data distribution abstractions and a locality-aware scheduling for NUMA systems and manycore processors.
Results show that data distribution abstractions enable programmers to distribute data without relying on fragile, low-level, and third-party methods, and that locality-aware task scheduling improves or maintains performance in comparison to work-stealing in return for minimal programmer effort

 

What will the future bring for your research topic?

I have plans to polish the grain graph workflow prototype to production quality and release it. Automating the optimization stages of workflow is also part of future work.
The defence committee and opponent were excited about the potential of grain graphs. I hope that industrial tool developers will be similarly excited and adopt grain graphs in their own tools.

 
 

What are your future plans? (like stay in Sweden, move, work, continue researching)

My future looks exciting. To put it pragmatically, I have realized the limit of my knowledge and want to push it further. There are so many things that I want to learn and work on! Practically, I am interested in pursuing an academic career in Sweden. Therefore my immediate plan is to take up a post-doctoral research position and start building a competitive academic profile.