llvm-devmtg-2022

Artifacts for 2022 US LLVM Developer Meeting presentation

Prerequisites

• Linux x86_64 (BOLT instrumentation limitation at the moment)
• ccache (to cache first-stage builds across configurations)
• perf: sudo apt-get install linux-tools-`uname -r`
  • Allow perf sampling:
  • Add kernel.perf_event_paranoid = -1 to /etc/sysctl.conf
  • sudo sysctl -p
• LLVM build prerequisites: CMake, Ninja, Python, etc
• 8 GB RAM or more
• 8 physical cores to run the full benchmark in reasonable time. 4 cores is the barest minimum but the benchmarking can take ~10h.

Launch instructions

git clone https://github.com/aaupov/llvm-devmtg-2022
cd llvm-devmtg-2022
bash -x ./driver.sh <XXX> |& tee log.txt

Where <XXX> is an arch name, one of BDW GLC GRT ZEN1. Note that your CPU needs to match the specified arch exactly by the core count and thread topology. Check the driver to make sure it's specified correctly, and make changes if needed.

Follow instructions at https://llvm.org/docs/Benchmarking.html#linux to find out SMT pairs that need to be disabled for the benchmark.

If you have a local up-to-date LLVM checkout, you can save time by providing the repo to the script:

LLVM_SOURCE=/path/to/llvm-project bash -x ./driver.sh <XXX> |& tee log.txt

Results

The results would be located in the temporary folder of the run (which is not deleted automatically), which is printed in the beginning of the script.

Naming scheme: $CONFIG_$ARCH_run.txt, e.g. BASELINE_GLC_run.txt or BOLT_PGO_GRT_run.txt