The roadrunner project and the importance of energy efficiency on the road to exascale computing

ABSTRACT

The cost of the energy required to run ultrascale supercomputers is becoming a large portion of the operating budgets of many facilities and it has the potential of limiting the scale of computers that can be deployed. The Roadrunner project was started as a development project aimed at finding a way to scale up applications but at a significantly more efficient energy usage than the current systems. Heterogeneous core types allow single thread performance to remain high while reducing the energy required for a given computation by eliminating the circuits and associated power that are not needed for the computation. This optimizes the energy cost per operation but puts a burden on the software to deal with heterogeneous core types.

The Roadrunner system was the first to reach a sustained Petaflop on the Linpack benchmark and it involved some interesting new Hardware but the bulk of the effort was in Software development including programming models and applications. Several applications were ported to the new structure with relatively little difficulty and the expected performance and energy efficiency improvements were attained.

This talk will cover an overview of the Roadrunner project, including the fundamental Cell BE building block and the software structure and methods that were included. The success of the energy efficiency improvement has lead to a broader view of the utility of heterogeneous computing in the Computationally Intensive Workload area.

INDEX TERMS

Primary Classification:
  C. Computer Systems Organization
  C.1 PROCESSOR ARCHITECTURES
      C.1.4 Parallel Architectures

Additional Classification:
  C. Computer Systems Organization
  C.5 COMPUTER SYSTEM IMPLEMENTATION
      C.5.0 General

  D. Software
  D.1 PROGRAMMING TECHNIQUES
      D.1.3 Concurrent Programming

General Terms:
Algorithms, Design, Measurement, Performance

Keywords:
cell broadband engine, exaflop, exascale computing, heterogeneous multicore architectures, petaflop