High-Performance Computing Can Accelerate Life Sciences Discoveries

By Bhanu Rekepalli, PhD 

December 31, 2014 | Inside the Box | In the late 2000s, supercomputers reached petascale: executing a quadrillion calculations per second. Early petascale systems used conventional processors that consumed massive amounts of electricity. To reduce energy consumption, in recent years pace-setting companies such as Intel, Nvidia, and AMD have released energy-efficient, high-performance chips known as accelerators or coprocessors.  Examples include Intel’s Many Integrated Core (MIC) architecture Xeon Phi, Nvidia’s Graphics Processing Unit (GPU) architecture Kepler, and AMD’s FirePro S10000 graphics card.

Scientific computing users appreciate these high performance computing (HPC) chips for their ability to speed up modeling and simulation, data analysis, and visualization. In addition, chip manufacturers are working on efficiently combining Central Processing Units (CPUs), coprocessors, and built-in logic for interfacing with other devices on the same chip: a system-on-a-chip (SoC). This SoC technology will bring massive computing power—equivalent to a previous generation supercomputer—to desktops. Affordable personal supercomputers will be a reality in the near future.

Life sciences in particular could benefit immensely from this massive growth in HPC processing power, but the industry is not yet ready for petascale resources. Most life sciences applications are developed for desktop machines or small clusters, and few efforts are underway to port or adapt the existing life sciences applications to supercomputers.

One such effort is the work done in my former lab at the University of Tennessee, where the most widely used life sciences informatics tools are optimized and scaled to tens of thousands of cores for the fastest supercomputers. These parallel bioinformatics tools can perform tens of millions of genomic sequence search analyses using the one hundred thousand cores on petascale supercomputers.

For example, using the Titan supercomputer at Oak Ridge National Laboratory, we were able to run NCBI BLAST searches of 150 million DNA sequences, representing an entire animal gut microbiome, against a database of 50 million nucleotide sequences, in less than 48 hours. Computations on this scale would take months on a typical, medium-size cluster. The University of Tennessee has also used its Kraken supercomputer to dramatically speed up other scientific applications. In one experiment, we used optimized parallel cheminformatics tools to screen millions of small molecules against cancer pathway proteins in search of new drug targets. Although this process would have taken years to yield potential targets on a cluster, it was completed in a matter of a few days on Kraken.

The high memory supercomputer Nautilus was even used in a recent high-profile study to understand the evolutionary relationships among birds. Nautilus resolved phylogeny at the chromosome level using the whole genomes of 48 species of modern birds, from every order of the class Neoaves.

Chip-manufacturing companies are also optimizing life sciences applications to run efficiently and rapidly on their respective hardware architectures. IBM is optimizing applications on their POWER processors for data analytics. Intel and Nvidia are optimizing applications on their coprocessors or accelerators.

Hybrid architectures with greater parallelism and processing power in a smaller energy footprint are changing the way computational scientists approach problems. This technology is still in its infancy, and much work is needed to fully realize the potential of powerful next-generation technologies.

Intel’s Xeon Phi coprocessor is based on x86 architecture, the Intel CPU architecture that has dominated supercomputing for the past few decades. This conventional foundation means that Phis can use established programming tools, such as C/C++, Fortran, OpenMP, OpenCL and math libraries, with which most of the scientific computing community is familiar.  Massive efforts are underway in Intel’s Parallel Computing Centers (IPCCs) at 20 different academic institutions in the Americas, Europe, and Asia, where experts are optimizing and scaling select scientific application codes on Xeon Phi architectures. However, fewer than 10% of these codes are for life sciences applications.

GPUs, meanwhile, have been introduced for manipulating computer graphics and images, and have become popular in scientific computing over the past decade due to their great ability to exploit a high degree of parallelism. Nvidia, which holds a major share of the GPU market, has a dozen or so life sciences applications optimized and scaled to run on their general purpose GPUs. These GPUs, however, have a stream processor architecture and can only be programmed with Nvidia’s proprietary CUDA parallel computing platform, OpenACC, and OpenCL framework for writing programs that run on hybrid architectures. Most of the life sciences community is not familiar with these approaches, with the result that few biologists can make efficient use of GPU architectures.

To address the challenges of managing the volume, velocity, and variety of big data on petascale machines, the culture and practices of life sciences application developers and bioinformaticians will need to change. They will have to work collectively alongside hardware developers to co-design applications that make the most of architectures and programming environments.

Meanwhile, a major mission of the US Department of Energy (DOE) is to deploy a system a thousand times faster than first petascale system. The coming exascale (executing a quintillion, or thousand quadrillion, calculations per second) system will likely be deployed in the early- to mid-2020s, according to HPC experts, and will be designed for even greater energy efficiency, interconnect and memory performance, extreme parallelism, programming and data management, algorithm reinvention, resilience, and scientific productivity.  This increase in computing power will transform the global economy and scientific computing world. Life sciences applications need to be rewritten at an algorithmic level to anticipate advances possible from future exascale architectures, while solving today’s life sciences problems at petascale.

Inside the Box is a regular Bio-IT World column written by members of the BioTeam, a Massachusetts-based high-performance consulting practice. Bhanu Rekepalli is a Senior Scientific Consultant and Principal Investigator at The BioTeam, Inc. He can be reached at bhanu@bioteam.net 

Further Reading 

Bhanu Rekepalli, Paul Giblock and Christopher Reardon. “PoPLAR: The Portal for Petascale Lifescience Applications and Research”, BMC Bioinformatics 2013, 14:s3.

Bhanu Rekepalli, Aaron Vose and Paul Giblock. “HSPp-BLAST: Highly Scalable Parallel PSI-BLAST for Very Large-scale Sequence Searches,” 3rd International Conference on Bioinformatics and Computational Biology (BICoB). Edited by Saeed F, Khokhar A, Al-Mubaid H. Las Vegas, NV; 2012:37-42

Erich Jarvis, et al. “Whole Genome Analyses Resolve the Early Branches in the Tree of Life of Modern Birds,” Science 2014, Vol. 346 no. 6215 pp.1320-1331