Skip to content

bwelton/diogenes

Repository files navigation

Diogenes

Diogenes is a performance tool to identify unnecessary/misplaced synchronizations and memory tranfers. Diogenes provides an estimate of potential benefit of fixing the unnecessary/misplaced operation.

This project is very much a work in progress and is very much in a research state (lots of dead code, difficult to build, lots of bugs, no polish, etc).

NOTE

There is an enormous bug fix patch in the works. In addition a few new features will be added soon. In the near future, work will begin on generalizing Diogenes to take it out of the PhD realm and into something usable by other (such as the ability to run on more systems other than LLNL's machines). If you are interested in running Diogenes on your machine, contact me at welton [at] cs.wisc.edu.

There is currently a slow effort to refactor this codebase into a more manageable state (including with performance imporvements such as by using page protection for memory tracing). This codebase is in the branch "redesign" and includes many of the components needed for autocorrection (memory transfer autocorrection may have bugs in the redesign branch). For those who intend to work on this in the future, you would be best served by looking at that branch.

Building and Installation

Due to the number of dependencies that this project has (Dyninst, Boost, Cuda, Libdwarf, Libelf, and others), it is highly recommended you build this package with the spack package manager (https://github.com/LLNL/spack). A premade spack repo package for this project is availible in spack/. Adding ./spack to ~/.spack/repos.yaml will allow you to build the deduplicator with the following commands:

cd SPACK_DIR ./bin/spack install cudadedup

Required build parameters if building manually (unwise)

The following parameters are required at CMake configure time if not set by environment variables.

  • -DCUDA_TOOLKIT_ROOT_DIR=Location of Cuda Toolkit (Optional on Cray Machines)
  • -DLIBCUDA_SO=Absolute path to libcuda.so, this should be the path to libcuda that applications are linked against (Optional on Cray Machines)
  • -DLIBCUDART_SO=Absolute path to libcudart.so (Optional on Cray Machines)
  • -DLIBELF_LIBFILE=Absolute path to Libelf.so (Optional on ORNL Cray Titan)
  • -DLIBELF_INCLUDE=Path to libelf include directory (Optional on ORNL Cray Titan)
  • -DLIBDWARF_LIBFILE=Absolute path to libdwarf.so, CANNOT be the static version of libdwarf (Optional on ORNL Cray Titan)
  • -DLIBDWARF_INCLUDE=Path to libdwarf include directory (Optional on ORNL Cray Titan)

Optional Build Parameters

The following parameters are not required but can reduce compilation time if set (by skipping compilation of python/dyninst)

  • -DDYNINST_ROOT=Path to the installation directory of dyninst (containing /lib and /include)
  • -DPYTHON_ROOT=Path to the installation directory of python 2.7
  • -DCMAKE_INSTALL_PREFIX=Installation Directory

Disclaimer

This software should be considered research grade right now, use at your own risk.