Ohio State Navbar

The Ohio State UniversityOffice of International Affairs

give a donation

Efficient asynchronous communication progress for MPI without dedicated resources

Research Scholar

Amit Ruhela, senior research engineer, Indian Institute of Technology Delhi (India)
Dhabaleswar K. Panda, faculty mentor

Background

  • Hometown: Kurukshetra, India
  • Degrees received: PhD in computer science, Indian Institute of Technology Delhi, India.

What is the issue or problem addresses in your research?

To reduce running time of MPI fueled scientific applications, asynchronous progress thread are used. We aim to improve their performance via the following goals:

  • Can we identify when async progress is required and when to activate it?
  • Can we minimize the CPU utilization of the async thread and maximize the CPU availability for application computation?
  • Can we avoid using any specialized hardware/software resources?
  • What key design changes are required and how to keep the simplicity of our design?

What methodology did you use in your research?

  • In-depth study of existing approaches.
  • Explain reasons why existing approaches are not generic and deliver subdue performance.
  • Propose a simple yet effective design that defies the commonly accepted trend.
  • Compared and evaluated the proposed design at both microbenchmark level and application level across three different MPI libraries and four distinct hardware architectures.
  • Implemented design in Mvapich2 library that is used by more than 2,925 organizations in 86 countries.

What are the purpose/rationale and implications of your research?

The fundamental purpose of our research is to optimize the performance of MPI libraries which powers several scientific applications. This in-turn would facilitate HPC applications to process a huge volume of data in a shorter time. Consequently, predictions made by these applications would be more accurate and timely. At a deeper level, our goal is to design a simple and smart technique that don't necessitate extra hardware and software resources and works flawlessly with leading mpi libraries.