Marvin Doyley, Research Assistant Professor at Thayer School of Engineering, is part of Keith Paulson’s alternative breast imaging team that is working on techniques that may one day allow earlier detection and classification of breast tumors. Early discovery and classification of these tumors will improve treatment options and also help avoid the many unnecessary biopsies now required to determine tumor type. Doyley uses an emerging technique called magnetic resonance elastography (MRE) in which a small amount of motion is applied to tissue. Since a tumor is many times stiffer than normal tissue, it deforms less than the surrounding tissue. The Thayer School research group has developed a technique where they use a computationally intensive numerical method to visualize the stiffness distribution within the breast, which allows them to differentiate between diseased and normal breast tissues. They make an initial guess of the stiffness distribution within the breast, calculate the motion of the tissue, and compare the calculated results to magnetic resonance imaging (MRI) results. They then use an iterative process, refining their last guess and repeating the calculation until the computed results match the MRI results.
Utilizing the resources of the Discovery cluster has allowed Doyley to reduce job runtimes from one week down to four hours. Utilizing 48 of Discovery’s 300 CPUs has sped up the research process, reports Doyley, and has allowed the team to work on more complex problems then in the past. In assessing the impact that more computational resources has had on his research, Doyley is quite clear about the impact the of Discovery: “Discovery gives us a lead. It allows us to be more competitive.”
Discovery is a collection of 101 computers, or nodes, with a total of 342 CPUs, 11 Terabytes of disk space, and 600 Gigabytes of memory. The combined power of these interconnected nodes allows for fast processing of programs. In addition, some Discovery nodes have special infiniband hardware that allows them to communicate up to five times faster than the normal Gigabit connections. The Discovery cluster was conceived and funded by Dr. Jason H. Moore, Director of Bioinformatics at DMS, to become a campus-wide coop cluster designed to serve the needs of Dartmouth researchers and students.
With this faster interconnect, users running parallel programs that run on multiple nodes are able to run their programs faster and more efficiently on large numbers of processors. For example, a physics application that uses message passing (MPI) was able to run about 35% faster on Infiniband.
Here are some benefits of using the Discovery cluster:
- Access to a 340 CPU cluster
- Full system administration support, including backups
- Services including programming, debugging, parallelization, and optimization support
- Full suite of standard compilers and research applications
- Access to inactive nodes
- High speed connections between some nodes
For more about Discovery, including information about courses, visit discovery.dartmouth.edu or contact John Wallace.
- Pete Schmitt (646-8109), Systems Administrator
- Susan Schwarz (646-1458), Training and Application Development
- John Wallace (646-1412), Outreach and Consulting