Optimizing the Performance of Streaming Numerical Kernels on the IBM Blue Gene/P PowerPC 450 Processor

​Streaming Kernels in Scientific Applications​

• Bottleneck in scientific codes

• KAUST applications:

• ​​​Explicit high-order methods for hyperbolic PDEs (Ketcheson)

• Inverse seismic imaging (Schuster)

• Navier-Stokes-Korteweg (Calo)

• Plasmoid simulations (Samtaney)

​​

​Challenges in Tuning for PowerPC 450 Processor

• Inhibition in the utilization of its SIMD and the dual-issue pipeline

• PowerPC 450 requires reordered instructions to complete earlier

Efficient SIMD Algorithm for the 3-point Stencil

• The 3-point stencil is a building block to many stencils

• Fully utilizes SIMD-like capabilities with no instruction waste

Python Code Synthesis and Modeling Framework

• Simplifies coding with a faster development-testing loop

• Automates out-of-order scheduling and cycle-acurate performance modeling

Modeling and Performance Within L1 Cache

• Optimization-enabling accurate modeling of the PowerPC 450 pipeline

27-point Stencil Results

• 1.72x speedup over the best published results for large size problems [1]

• 2.16x speedup over optimized C codes for domain size fitting in L1 cache

References

1. 1.​K. Datta, ``Auto-tuning Stencial Codes for Cache-Based Multicore Platforms", PhD Thesis, EECS Department, University of California, Berkeley, December 2009

2. 2.T. Malas, A. Ahmadia, J. Brown, J. Gunnels, and D. Keyes, ``Optimizing the Performance of Streaming Numerical Kernels on the IBM Blue Gene/P PowerPC 450 Processor", Conditionally accepted, IJHPCA Journal