Scientific Computing Group Seminars - Detail View

Speaker: L Barba

Affiliation: Boston University

Talk Title: The Need for Speed: Fast Algorithms, Fast Hardware (Is Not Always Equal to Fast Code)

Invited by: George Karniadakis

Time: April 02 2010 11 a.m.

Location: 182 George Street, Room 110

Abstract:

Scientific computing has an insatiable hunger for speed and for computing capacity. As we attempt to solve more and more complex problems on computers -- predict global climates, unravel the mysteries of life by simulating protein interactions, and so on -- we need faster and faster computing. Moore's law was interpreted to mean a doubling of speed every 18 months, so maybe some will be tempted to just wait for computers to get fast enough. But the importance of clever (fast) algorithms cannot be underestimated. For example, if a problem has one million unknowns, and there is a solution algorithm which requires in the order of N-squared operations, it would take 16 generations of Moore's law (24 years) before this solution is comparable to a fast algorithm requiring in the order of N operations. And recently, a new wave of "accelerated computing" is taking flight. The impact of hardware acceleration, for example, using graphics processors to offload arithmetically intensive portions of a code, is now unmistakable. I will discuss research efforts in my group involving both fast algorithms -- in particular the fast multipole method, and fast radial basis function methods -- and hardware acceleration using graphics processors (known as GPUs). The combination should be explosive, but achieving the speed in actual code is not straightforward. Most times, an algorithm will require an honest re-think before we can tap the speed of the accelerator hardware. I will give examples using the fast multipole algorithm. I will also motivate by briefly presenting the variety of applications where our work can have an impact.