New Technology I
           
Event Type Start Time End Time Rm # Chair  

 

Masterworks 1:30PM 2:15PM 16-18 Norm Morse (Lawrence Livermore National Laboratory)
 
Title:
Trends in High-Performance Computer Architecture (It's the Bandwidth)
  Speakers/Presenter:
Bill Dally (Stanford University) New

 

Masterworks 2:15PM 3:00PM 16-18 Norm Morse (Lawrence Livermore National Laboratory)
 
Title:
The Future of Supercomputing Software
  Speakers/Presenter:
Susan Graham (University of California, Berkeley)
             

 

     
  Session: New Technology I
  Title: Trends in High-Performance Computer Architecture (It's the Bandwidth)
  Chair: Norm Morse (Lawrence Livermore National Laboratory)
  Time: Wednesday, November 19, 1:30PM - 2:15PM
  Rm #: 16-18
  Speaker(s)/Author(s):  
  Bill Dally (Stanford University) New
   
  Description:
  Modern scientific computations are bandwidth, not arithmetic, limited. Today, a 1GHz 64-bit FPU has an area of less than 0.5mm^2 making the raw cost of a GFLOPS of arithmetic less than 50 cents. This cheap arithmetic makes peak FLOPS inexpensive and irrelevant. Sustained FLOPS, which are relevant (and more expensive), are largely determined by local and system-wide memory (DRAM) bandwidth. In addition to providing adequate raw bandwidth, an efficient system must provide a latency-hiding mechanism to sustain this bandwidth while waiting on main memory (or network) latency, and must provide high bandwidth on single-word references (gathers and scatters). This talk will discuss how the architecture of high-performance computers - both commodity and custom - address local and global bandwidth issues. These machines use different methods to provide raw bandwidth, hide latency, and support irregular accesses. Emerging techniques, including streams and processor-in-memory, hold the promise of enhancing locality to reduce bandwidth demand. A machine with a global address space and high random-access bandwidth simplifies application software by reducing the importance of placement (in time and space)- allowing the programmer to focus on the fundamental issues of parallelism and load balance.
  Link: --
   

 

     
  Session: New Technology I
  Title: The Future of Supercomputing Software
  Chair: Norm Morse (Lawrence Livermore National Laboratory)
  Time: Wednesday, November 19, 2:15PM - 3:00PM
  Rm #: 16-18
  Speaker(s)/Author(s):  
  Susan Graham (University of California, Berkeley)
   
  Description:
  Although supercomputing hardware deserves much of the credit for increasing speed of computations, it is not the whole story. Indeed, this hardware is essentially useless without good software, from the low-level operating systems to optimized applications solving real problems. Moreover, advances in algorithms and software multiply the effects of improved hardware. For example, compiler register allocation and cache optimizations can improve application performance by 50%, the equivalent of moving to the next generation of processor. Yet many reports, most recently the National Academy of Science study "The Future of Supercomputing" (http://www7.nationalacademies.org/cstb/project_supercomputing.html), have noted that research in supercomputing software is severely underfunded. In this talk, we look at the major challenges for supercomputer software, including supporting existing applications and moving toward revolutionary system designs. We will also look at some promising techniques for ensuring that tomorrow's software is better (faster, more reliable, more maintainable) than today's.
  Link: --