The Week in HPC Research
The top research stories of the week have been hand-selected from major science centers, prominent journals and leading conference proceedings. Here’s another diverse set of items, including whole brain simulation; a look at High Performance Linpack; the coming GPGPU cloud paradigm; heterogenous GPU programming; and a comparison of accelerator-based servers.
Brain Simulation Project
The Human Brain Project, one of the most ambitious projects of its kind, has just been awarded half-a-million Euros over a 10-year timeframe. The European Commission funded the innovative program as part of its Future and Emerging Technologies (FET) flagship program. Led by Henry Markram, a neuroscientist at the Swiss Federal Institute of Technology in Lausanne, the project aims to reconstruct the brain piece-by-piece, using cutting-edge supercomputing resources.
According to the announcement:
As a result of this initiative, in neuroscience and neuroinformatics the brain simulation will collect and integrate experimental data, identifying and filling gaps in our knowledge. In medicine, the project’s results will facilitate better diagnosis, combined with disease and drug simulation. In computing, new techniques of interactive supercomputing, driven by the needs of brain simulation, will impact a range of industries, while devices and systems, modelled after the brain, will overcome fundamental limits on the energy-efficiency, reliability and programmability of current technologies, clearing the road for systems with brain-like intelligence.
The “Human Brain Project” is on track to become the world’s largest experimental facility for developing the most detailed model of the brain. The research will increase our understanding of how the human brain works, which has countless implications for technology and medicine, from personalized medical treatments to artificial intelligence breakthroughs.
Researchers are divided over the news. Detractors say it’s an impossible endeavor at our current stage of computational development to model the brain’s 86 billion neurons. To make it really interesting will mean capturing the brain’s actual creative potential and intelligence, otherwise it will just be a big computer.
An Investigation into High Performance Linpack
A research item in the Proceedings of 2012 2nd IEEE International Conference on Parallel Distributed and Grid Computing, which took place Dec. 6-8, 2012, presents an analysis of process distribution in HPC cluster using High Performance Linpack.
The authors, a group of computer scientists from the Raja Ramanna Centre for Advanced Technology in Indore, India Computing acknowledge the fact that scientific endeavors increasingly rely on parallel programming techniques running on High Performance Computing Clusters (HPCC).
When it comes to measuring cluster performance, there are multiple factors to take into account. “Memory, interconnect bandwidth, number of cores per processor/ node and job complexity are the major parameters which affect and govern the peak computing power delivered by HPCC,” they write.
The paper describes the researchers’ experiments with High Performance Linpack (HPL). They use the benchmark to analyze the effect of job distribution among single processors versus distributed processors. They’re also investigating the effect of the system interconnect on job performance. The work centers on an InfiniBand-connected HPC cluster.
The GPGPU Cloud Paradigm
The increasing prevalence of hybrid HPC systems that use coprocessors like GPUs to improve performance has implications to HPC cloud. In a new research paper [PDF], a team of computer scientists from the College of Computer Science and Technology at Jilin University in Changchun, China, explores the idea of GPGPU cloud as a paradigm for general purpose computing. Their work appears in the February 2013 issue of the Tsinghua Science and Technology Journal.
The authors start with the premise that the “Kepler General Purpose GPU (GPGPU) architecture was developed to directly support GPU virtualization and make GPGPU cloud computing more broadly applicable by providing general purpose computing capability in the form of on-demand virtual resources.”
To test their theories, they developed a baseline GPGPU cloud system outfitted with Kepler GPUs. The system is comprised of a cloud layer, a server layer, and a GPGPU layer, and the paper further describes “the hardware features, task features, scheduling mechanism, and execution mechanism of each layer.” The work aims to uncover hardware potential while also improving task performance. In identifying the advantages to general-purpose computing on a GPGPU cloud, the authors show themselves to be on the forefront of an emerging paradigm.
Heterogeneous Computing on GPU Clusters
A group of scientists from the University of Minnesota and University of Colorado Boulder have contributed to a recently-published book, GPU Solutions to Multi-scale Problems in Science and Engineering. Their chapter, titled High Throughput Heterogeneous Computing and Interactive Visualization on a Desktop Supercomputer, examines some of the computational improvements that have resulted from the GPU accelerator movement. Their test system, a “desktop supercomputer,” was constructed for less than $2,500 using commodity parts, including a Tesla C1060 card and a GeForce GTX 295 card. The GPU cluster runs on Linux, and employs CUDA, MPI and other software as needed.
The authors make some interesting observations, including the following:
MPI is used not only for distributing and/or transferring the computing loads among the GPU devices, but also for controlling the process of visualization. Several applications of heterogeneous computing have been successfully run on this desktop. Calculation of long-ranged forces in the n-body problem with fast multi-pole method can consume more than 85 % of the cycles and generate 480 GFLOPS of throughput. Mixed programming of CUDA-based C and Matlab has facilitated interactive visualization during simulations.
They explain that what sets their work apart from other published research is their use of multiple GPU devices on one desktop, employed by multiple users for various types of applications at the same time. They state that they have extended GPU acceleration from the single program multiple data paradigm to the multiple program multiple data paradigm, and claim “test runs have shown that running multiple applications on one GPU device or running one application across multiple GPU devices can be done as conveniently as on traditional CPUs.”
Accelerators Compared for Energy Efficiency
The entire book, GPU Solutions to Multi-scale Problems in Science and Engineering, is quite fascinating. Another chapter written by University of Houston’s Lennart Johnsson explores the energy efficiency of accelerated HPC servers.
Johnsson traces the evolution of mass market, specialized processors, including the Cell Broadband Engine (CBE) and graphics processors. She notes that GPUs, in particular, have received significant attention. The addition of hardware support for double-precision floating-point arithmetic, introduced three years ago, was key to this signification uptick in adoption, as was the recent support of Error Correcting Code.
To analyze the feasibility of deploying accelerated clusters, PRACE (the Partnership for Advanced Computing in Europe) performed a study, investigating three types of accelerators, the CBE, GPUs and ClearSpeed. The study assessed several metrics, including performance, efficiency, power efficiency for double-precision arithmetic and programmer productivity.
In this chapter, titled “Efficiency, Energy Efficiency and Programming of Accelerated HPC Servers: Highlights of PRACE Studies,” Johnsson presents and analyzes some of the results from those experiments. She observes that the “GPU performed surprisingly significantly better than the CPU on the sparse matrix-vector multiplication on which the ClearSpeed performed surprisingly poorly. For matrix-multiplication, HPL and FFT the ClearSpeed accelerator was by far the most energy efficient device.”
Inaugural HPC Award Winners
The Department of Energy’s National Energy Research Scientific Computing Center (NERSC) unveiled the winners of their inaugural High Performance Computing (HPC) Achievement Awards. The announcement was made at the annual NERSC User Group meeting at the Lawrence Berkeley National Laboratory (Berkeley Lab).
All NERSC users, the awardees were selected for their innovative use of HPC resources to help solve major computational or humanitarian challenges. Two early career awards were also presented.
NERSC Director Sudip Dosanjh stated that “High performance computing is changing how science is being done, and facilitating breakthroughs that would have been impossible a decade ago. The 2013 NERSC Achievement Award winners highlight some of the ways this trend is expanding our fundamental understanding of science, and how we can use this knowledge to benefit humanity.”