NERSCPowering Scientific Discovery Since 1974

Green Computing


March 1, 2008

NERSC and other Berkeley Lab  researchers are taking on energy efficiency research that aims to influence the  computing industry in designing and building computer and storage technologies  that will benefit the scientists, consumers  and the environment.  

Thanks to the Laboratory Directed  Research and Development Program  (LDRD) at Berkeley Lab, researchers are  exploring subjects in computer architectures, algorithms and mass storage system designs to improve energy efficiency  of scientific computations. The LDRD program provides special funding for promis- ing research projects, and this set of related LDRD projects includes researchers  from NERSC and the Computational  Research Division (CRD) at Berkeley Lab. 

Pursuing energy efficient computing  research at NERSC makes sense, both to  save energy and to shift NERSC resources  away from energy costs and towards systems and services that directly benefit the scientific community. With over 3,000 users  and an insatiable demand for NERSC  computing and storage facilities, results  from the LDRD projects could indirectly  benefit scientists across scientific disciplines that rely on computing. This includes cosmology, climate, life sciences, accelerator physics, fusion, computer science  and material science.  

Kathy Yelick, the NERSC Director,  explains the importance of this work.  “Power is the most important problem in  computing today, not just at the high end,  but from hand-held devices and laptops  to data centers and computing centers  like NERSC. Power density within chips  has forced the entire processor industry  to put multiple cores on a chip, and within  centers the total system power is a major  component of cost and availability.”  

She describes this set of LDRDs as a  “multi-faceted attack” on the problem,  starting with a blank slate on the architecture end, and rethinking algorithms, applications and software to make use of energy- efficient hardware. The first goal is to do  more science with less energy, and the  second is to enable the next generation of  exascale computing systems, which require  technological breakthroughs to address  the power issues at such extreme scales. 

One of the projects takes a vertical  slice through the problem space, looking  at a single application domain and considering alternative algorithms as well as  architectures for solving the problem.  Climate modeling is the target application,  selected because of its significance to  science and the general public, and  because it requires millions of CPU computing hours to explore various climate  scenarios and the possible impacts of  changes in policy or alternate fuel  sources. The other LDRD projects take a  broader look at specific aspects of the  problem, including energy-efficient computing components based on multicore  technology, energy efficient storage systems, and application characterization  that explores the ability of various key  algorithms to adapt to energy-efficient  hardware. 

Here are the four LDRD projects: 

Climate Modeling System 

The development of multicore chips  has been the computer industry’s solution to keep power consumption in check. But  some of the current approaches to adding  more complex cores per chip would eventually hit a performance plateau. John  Shalf, Lenny Oliker and Michael Wehner  are investigating an alternative to using  conventional microprocessors in designing energy-efficient supercomputers that  employ more aggressive use of parallelism and design techniques from the  consumer electronics industry to more  closely tailor the chip design to the needs  of scientific applications. They anticipate  this approach could achieve 100 times or  more improvement in power efficiency  and effective performance over business  as usual. 

Their Climate Simulator Project will  build a prototype system using embedded  processors — low-power chips commonly  found in consumer electronics devices  such as cell phones and portable music  players — and tailor its performance to  provide optimal power efficiency for climate modeling problems. The group has  adopted battery-powered designs in consumer electronics that are very sensitive  to power consumption and cost. These  embedded chips are less powerful and  less power hungry than conventional  microprocessors in supercomputers these  days, and can be more easily customized  to run specific applications. A computer  built with thousands of these embedded  chips could extract the most energy-efficient performance that also is capable of  tackling complex scientific problems. The  approach not only promises to be more  power-efficient than the conventional path  forward in HPC — it also promises to be  more cost-effective.  

The Climate Simulator team is working  with Tensilica, an embedded processor  design firm, as well as David Randall, a  professor in the Atmospheric Sciences  Department at Colorado State University  and a NERSC user. Randall’s climate  modeling code, developed under DOE’s  SciDAC program, is a new breed of cli- mate modeling code that is capable of  expressing enough parallelism to run kilometer-scale simulations 1000 times faster  than real time on machines envisioned by  the Climate Simulator research team. In  order to enable dramatic changes in  power efficiency, codes such as Randall’s  must expose orders of magnitude more parallelism than the current climate mod- els. Employing simpler processors that  are designed for parallel throughput  rather than serial performance will  enable substantial power efficiency  gains. 

“We want to find compelling solutions  to scientific problems that need petascale machines,” Shalf said. “The use of  these power-efficient cores will help us  achieve those goals.” 

Manycore Chips 

The development of multicore chips  represents the most significant shift in  microprocessor engineering in several  decades, and it opens up opportunities  for exploring innovative designs for high- performance computers.  

Jonathan Carter, head of the User  Services Group at NERSC, is leading the  project to explore a wide range of multi- core computer architectures and how  efficiently those systems can perform on  challenging scientific codes. The project,  “Enhancing the Effectiveness of  Manycore Chip Technologies for High- End Computing,” also includes collabora- tors Lenny Oliker and John Shalf. 

Future supercomputers will likely be  built with chips containing an increasing  number of cores. Multicore chip designs  vary, however. They include heterogeneous designs, such as the Cell processor,  developed by IBM, Sony and Toshiba;  graphics processing units (GPUs); and  processors for the embedded market.  There are also homogeneous designs,  such as microprocessors by Intel and  Advanced Micro Devices, the world’s two largest chip makers. In many cases, multicore technologies offer higher absolute  performance and more energy-efficient  computation. 

“This LDRD project provides a  breadth of architecture coverage to our  whole ultra-efficient research thrust. We  want to identify candidate algorithms that  map well to multicore technologies, and  document the steps needed to re-engi- neer programs to take advantage of  these architectures,” Carter said. “In  addition, perhaps there are design elements in multicore chips that we can  influence to help design a better high- performance system.”

Mass Storage 

Led by CRD scientists Ekow Otoo and  Doron Rotem, the “Energy Smart Disk- Based Mass Storage System” project sets  out to investigate energy-efficient disk  storage configurations that also provide  quick access to massive amounts of data.  

Today’s storage systems in data centers  use thousands of continuously spinning  disk drives. These disk drives and the necessary cooling components use a substantial fraction of the total energy consumed  by the data center. As the need for reliable  long-term storage of data grows, so will the  associated energy costs.  

Otoo and Rotem have set out to explore  new configurations that divide the disks into  active and passive groups. The active  group contains continuously spinning disks and acts as a cache for most frequently  accessed data. The disks in the passive  group would power down after a period of  inactivity. Besides looking at optimal disk  configurations and file placement algo- rithms, the researchers will also develop  simulation models for analyzing energy use.  

Benchmarking for Dwarfs 

A project led by Erich Strohmaier proposes to develop a test bed for bench- marking of key algorithms that will be crucial for designing software and computers  that use processors with many cores on  each chip. This project is conducted with  domain experts from CRD, NERSC, and  UC Berkeley.  

From desktop PCs to supercomputers,  systems built with hundreds of cores or more are likely to hit the market as early  as the beginning of the next decade. The  scientific community and the computer  industry need to figure out how to make  efficient use of these more powerful  machines. This will be especially impor- tant for high end users, as they will face  systems with large differences in intercon- nect properties at different levels of the  system architecture hierarchy.  

The project, “Reference Benchmarks for  the Dwarfs,” will devise ways to use a set  of algorithms to gauge the performance of  systems from personal computers to high- performance systems. The algorithms are  known as dwarfs; each dwarf represents a  class of algorithms with similar properties  and behavior. The 13 dwarfs chosen for  the research include algorithms important  for the scientific community.

About NERSC and Berkeley Lab
The National Energy Research Scientific Computing Center (NERSC) is a U.S. Department of Energy Office of Science User Facility that serves as the primary high-performance computing center for scientific research sponsored by the Office of Science. Located at Lawrence Berkeley National Laboratory, the NERSC Center serves more than 7,000 scientists at national laboratories and universities researching a wide range of problems in combustion, climate modeling, fusion energy, materials science, physics, chemistry, computational biology, and other disciplines. Berkeley Lab is a DOE national laboratory located in Berkeley, California. It conducts unclassified scientific research and is managed by the University of California for the U.S. Department of Energy. »Learn more about computing sciences at Berkeley Lab.