MGHPCC Awards Seed Grants to Institute Members
Three research projects involving several Hariri Institute members have recently received “seed grants” for projects related to the Massachusetts Green High Performance Computing Center (MGHPCC). MGHPCC is a groundbreaking collaboration of five of the state’s most research-intensive universities—Boston University, MIT, Harvard, Northeastern University, and the University of Massachusetts—along with both state government and private industry, in the most significant collaboration among government, industry and public and private universities in the history of the Commonwealth, and the first facility in the nation of its kind.
As part of the launch of MGHPCC, a competitive award process selected several seed projects to begin work on research using the kinds of high-performance systems to be available at the Center. Three of those projects involved Boston University faculty members, combining support from the seed grant program, the Provost’s Office, and the Hariri Institute. The three BU projects are:
Designing Green Software for High Performance Computing Clusters
Ayse K. Coskun (BU), Gunar Schirner (NE), Martin C. Herbordt (BU)
Energy efficiency is one of the central societal and technical issues today, and has become yet more critical with increased focus on the environment and the world’s climate. Power consumption is also the major limit in improving computing performance. Large computing clusters already have prohibitive operational and cooling costs. In addition, further increases in processor power densities are not viable as they breach the reliable operating temperatures of computing hardware. These limitations on power consumption and cooling are bringing to an end the historic performance scaling in high performance computing (HPC) centers, which has so far enabled tremendous advancements in solving complex scientific problems.
This project’s research goal is to design systematic, inexpensive methods to optimize the application software for increasing the energy efficiency. HPC software has been traditionally optimized for performance, and little has been done for generating better software for energy efficient operation. As part of the research plan, the project will demonstrate proof-of-concept of novel “green software” techniques on real-life HPC applications.
Tree and Multipole Algorithms for Exascale Computing
Lorena Barba (BU), Cris Cecka (Harvard), Hans Johnston (University of Massachusetts, Amherst)
This project is a collaboration among investigators in BU, Harvard and UMass aimed at creating an open and high-performance software infrastructure for a family of hierarchical N-body algorithms. “N- body” is the name given to any problem that involves a system where each object depends on the state of every other object in the system. The classic example in mechanics is gravitational interaction, but the situation also appears in the interactions of atoms or ions and in discrete representations of the equations for acoustics, electromagnetics and fluid dynamics. Algorithms to solve this problem that use hierarchical groupings of the objects in the system (or points in the discretization) are often fast, and scale very well in highly parallel computers. In fact, some believe that these algorithms will be better suited to the increasing numbers of cores in high- performance computers, and more likely to reach exascale. What is lacking is a concerted effort to create high-performance software offering the power of these algorithms to the wider scientific community.
The team funded by this seed grant will develop software infrastructure for hierarchical N-body algorithms aimed at the top computing systems, such as Blue Gene (for which the support of IBM as an industrial partner will be instrumental) and GPUs. The team will also develop capability to solve diverse problems, including biophysics, acoustics and fluid dynamics.
Clinical Elastic High Performance Computing
Jonathan Appavoo (BU), Dr. Ellen Grant (Children’s Hospital and Harvard University)
One of the challenges in analyzing radiology data by computer is that it typically takes days to get the results, because the processing is computationally intensive. In a clinical setting, doctors and patients would benefit from getting results in seconds. To make this happen, doctors need to call upon massive computational power instantaneously, and they only use it for a short period of time. This kind of bursty use is common in cloud computing, but conventional cloud computing systems don’t have the high-speed interconnection between computers needed to process this kind of data. And conventional high-performance computing systems don’t have the interactive ability that doctors need.
In this project, Dr. Grant and Prof. Appavoo will investigate putting those pieces together to create systems for on-demand radiological analysis. This research provides Prof. Appavoo a set of real applications to explore his model and mechanisms for ‘Interactive High-Performance Computing’, and gives Dr. Grant the systems expertise and computational resources to evaluate how well an interactive approach can benefit doctors and patients.