Advanced Scientific Computing

Linear Algebra and QCDOC on Top 500
A. Dubinsky, Y. Gao, C. Jung, R. Bennett, D. Stampf, and Y. Deng

We have ported to the QCDOC supercomputer with up to 8K computing nodes the basic benchmarks from the NPB suite, representing typical Linear Algebra algorithms: BT, LU, SP and MG. LU is a regular-sparse, block (5x5) lower and upper triangular system solver. SP computes the solution of multiple, independent systems of non-diagonally dominant, scalar pentadiagonal equations. BT performs solutions of multiple, independent systems of block tridiagonal equations with a 5x5 block size. MG performs simple multigrid calculations and long-distance communications.

We are examining its strong scalability by using increased numbers of processors to solve fixed-sized problems. We have also investigated weak scaling by studying larger problem sizes. We are testing the impact of the QCDOC 6D-torus interconnect design and distributed memory organization by comparing benchmark results from QCDOC to the results from other supercomputer machines with different structures, such as BlueGene (IBM), and the Beowulf clusters, such as Stony Brook Galaxy computer.

Although QCDOC was designed as a special-purpose machine for Quantum Chromodynamics (QCD) problems, it has shown strong performance and scalability on some applications requiring solution of linear algebra algorithms, and we continue to improve the scalability and increase the system size. An 8K node version has reached 4.61 Tflops at 70.3% efficiency and it would have ranked #109 on the June 2006 Top500 Listing. This performance result has been submitted to the Top500 for the November 2006 Listing and we expect to be listed around #170.
 

Last Modified: January 31, 2008
Please forward all questions about this site to: Claire Lamberti