gem5 Runahead

Implementing the HPCA'20 paper Precise Runahead Execution (PRE) in gem5

Architecture Overview

Usage

Compile

• IMPORTANT: install SCons 4.3.0 (higher version will fail!)

  pip3 install scons==4.3.0

• Compile gem5

  cd gem5
  
  time scons build/RISCV/gem5.opt -j`nproc`
  # 6min51s on my 32-thread server

New Params for PRE

All new params are defined under BaseO3CPU, as listed below:

Name	Description	Default
enablePRE	Enable Precise Runahead Execution	False
enablePREBranch	Allow branch instructions in PRE	False
enablePREEarlyRecycle	Enable register early recycle in PRE	False
numPRDQEntries	Number of precise register deallocation queue entries	192
numSSTEntries	Number of stalling slice table entries	128
enableMJ	Enable my journal	False

Run with Standard Config

I provide an O3CPU config examples/three_level_o3.py that resembles the model in the PRE paper.

• Run with PRE enabled

  build/RISCV/gem5.opt --outdir m5out-pre examples/three_level_o3.py \
      -p system.cpu.enablePRE=True \
      tests/test-progs/hello/bin/riscv/linux/hello
  # Hello world!

• Run without PRE

  build/RISCV/gem5.opt --outdir m5out-non examples/three_level_o3.py \
      tests/test-progs/hello/bin/riscv/linux/hello
  # Hello world!

• Print statistics

  # PRE
  cat m5out-pre/stats.txt | grep numCycles
  # system.cpu.numCycles  52698  # Number of cpu cycles simulated (Cycle)
  
  # No PRE
  cat m5out-non/stats.txt | grep numCycles
  # system.cpu.numCycles  52828  # Number of cpu cycles simulated (Cycle)

How to Compile gem5 Faster

It's been proved that Clang-16 is faster than GCC-9 in more cases, so the way to compile gem5 faster is to use Clang!

• Get LLVM

  # Get LLVM 16.0.6 (stable)
  wget https://github.com/llvm/llvm-project/releases/download/llvmorg-16.0.6/llvm-project-16.0.6.src.tar.xz
  
  # Uncompress
  tar xf llvm-project-16.0.6.src.tar.xz
  
  # Rename
  mv llvm-project-16.0.6.src llvm-project

• Compile and install Clang and lld (a linker 10x faster than GNU ld)

  cd llvm-project
  
  # Note: change CMAKE_INSTALL_PREFIX if needed
  cmake -G Ninja -S llvm -B build          \
        -DLLVM_ENABLE_PROJECTS='clang;lld' \
        -DLLVM_INSTALL_UTILS=ON            \
        -DCMAKE_INSTALL_PREFIX=/usr/local  \
        -DCMAKE_BUILD_TYPE=Release
  
  time ninja -C build install
  # 13min43s on my 32-thread server

• OPTIONAL: Install PyPy (a python interpreter with JIT optimization)

  # Get PyPy 3.10
  wget https://downloads.python.org/pypy/pypy3.10-v7.3.12-linux64.tar.bz2
  
  # Uncompress
  tar xf pypy3.10-v7.3.12-linux64.tar.bz2
  
  # A safe way to use PyPy without overwriting your CPython
  export PATH=`pwd`/pypy3.10-v7.3.12-linux64/bin:$PATH
  export LD_LIBRARY_PATH=`pwd`/pypy3.10-v7.3.12-linux64/lib:$LD_LIBRARY_PATH
  
  # Install pip (note that `python3' now refers to pypy)
  python3 -m ensurepip
  
  # Install SCons again (cannot use CPython's SCons)
  pip3 install scons==4.3.0

• Compile gem5

  cd gem5
  
  time CXX=clang++ scons build/RISCV/gem5.opt -j`nproc` --linker=lld
  # 3min56s on my 32-thread server

• gem5 compile time comparison

  Toolchain	Time (s)	Ratio
  GCC 9.4.0	411	100%
  Clang 16.0.6	236	57%
  Clang 16.0.6 + PyPy	243	59%

Presentation

Implementation and Analysis of the Runahead-based Hardware Data Prefetching
Download: runahead_implement_analysis.pdf