Branching

High performance library and command line tools for analysing the performance of branch predictors

The simulator and implementations are optimised for performance; single-threaded it can process around 6.5GB/s of traces, and the analysis program automatically parallelises experiments - note that this will use all available processing power, use RAYON_NUM_THREADS=N to limit parallelism to N threads.

Provided Strategies

• Always taken
• Two-bit
• GShare, with variable history bits
• Profiled static (experimental strategy targeting single-program embedded devices)

Crate Structure

There are two library crates and two executable crates.

• Branchlib is a library implementing the core branch prediction functionality, including...
  • Traits for prediction strategies
  • Simulation harnesses
  • Implementations for common strategies
  • Input format parsing
• Branchcli contains some common command line parsing definitions for the simulation and analysis binaries
• Branchsim is a command line tool which can be used to simulate a particular strategy on a given input file
• Branchanalyse is a command line tool which can run many simulators on many files and return results as CSV data