Parabix: Scalable Performance for Streaming Text Processing

• Processor clock-speeds stopped increasing in 2005
• Processor advances since 2005 have focused on parallel processing capabilities
  • SIMD parallelism
  • multicore parallelism
  • hyperthreading

• Performance of traditional software tools largely does not benefit from these advances.

• Considerable programmer effort is often required to build applications that can exploit parallelism.

The Parabix Approach

• Design of a compiler framework that automatically scales to take advantage of SIMD and multicore parallel processing resources.
• Employs algorithms based on bitwise data parallelism
• Current focus: streaming text processing applications
• Showcase application: icgrep
  • Feature complete for pure regular expressions
  • Extensive support for Unicode (best available for greps?)
  • Dramatic performance advantages in many cases

Parabix Compiler Technology

• Many Small Compilers
  • Pablo
  • Character class compilers for code units
  • Unicode hierarchy compiler
  • Unicode property compiler
  • Regular expression compiler
  • Kernels
  • Pipeline compilers
    • Sequential
    • Segment-Pipeline Parallel
• Future work: How can these all be integrated into an easier-to-understand/use framework
  • Focus on grammar compilers

LLVM for Parabix

• LLVM is useful for the Parabix framework to handle core code generation
  • Back-end support for many processors
• Several issues with LLVM for Parabix
  • IDISA overrides are used for critical Parabix performance issues
  • LLVM does not handle vectors of i1, i2, or i4 well
  • LLVM bitcast model does not work well for vectors
• Future work: Systematic Redesign of LLVM for Parabix-style Operations
  • Redesign the type legalization to support SWAR
  • Eliminate bitcast through memory and GEP into vectors