Transactional Memory and Memory Allocators Interaction Study

This repository contains the tools used in our PPoPP'15 paper Performance Implications of Dynamic Memory Allocators on Transactional Memory Systems. It can be used to reproduce our results and, with the necessary changes, investigate the behavior of new allocators and/or (S)TMs.

The repository contains the following open-source software:

• Allocators
  • Hoard (version 3.10) - See http://www.hoard.org
  • GPerftools (version 2.1) - See https://code.google.com/p/gperftools
  • Intel TBB (version 4.1) - See https://www.threadingbuildingblocks.org
• Benchmarks
  • Synthetic programs (threadtest and intset)
  • STAMP (Community version) - See https://code.google.com/p/tm-benchmarks
• TM
  • TinySTM (version (1.0.4) - See http://tmware.org/tinystm
• Tools
  • malloc_count - Adpated from http://panthema.net/2013/malloc_count/
  • PAPI 5.2 - See http://icl.cs.utk.edu/papi/

Dependences

The scripts used to compile, run, and post-process the data rely on some linux-based tools. In general they should be available in a typical distro, but we are listing here the main ones. In case some of the scripts finish with a weird error message, this could be of help.

• Gnuplot
• R Project - See http://www.r-project.org

You should also check the dependences of each of the tools (allocators, STM, etc.) before compiling them.

Code Organization

The code is organized in the following directory structure:

• allocators
• benchmarks
  • memory
    • threadtest
      • results-paper
      • results-smallrun
  • stamp
    • data
    • results-paper
    • results-smallrun
    • scripts
    • trunk
• tm
  • tinySTM-1.0.4
• tools
  • malloc_profile
  • papi-5.2.0

The top directories contain the allocators, benchmarks, tm library and tools needed to run the experiments. Only the source code is provided. Binaries are provided as part of our artifact only.

Quick Start

The first thing you should do is compile the packages needed to execute the experiments. This includes the allocators, tools, TM libraries and, finally, benchmarks. We provide a simple bash script to compile each of them (please enter their respective directories before executing the commands):

• Allocators: ./gen-allocators.sh

• Tools (needed if cache miss data is required): ./gen-tools.sh

• Threadtest: make

• STAMP and Intset: ./compile-all.sh

If all succeeds you are ready to run the experiments. Most of the benchmarks have a script to do it. You might want to look in the benchmarks/stamp/scripts/ directory. The main script is execute.sh, followed by gen-data.sh and plot-graph.sh. They currently have a very limited but somewhat useful command-line help system. For instance, to execute the yada application 5 times you should run:

./execute.sh -a "yada" -n 5

This will run the application with the default STM library (tinySTM), with four allocators (glibc, hoard, tbbmalloc, and tcmalloc), 1, 2, 4, and 8 threads, five times each. Try using ./execute.sh -h for some help. The execution script will create a subdirectory named results (default) in the stamp directory to save all runs.

To post-process the data and generate the chart for the previous executions you could try the following:

./plot-graph.sh -d stamp -g bar-stamp.gnu -a "yada"

It will store the tables and chart inside a directory named stamp (flag -d). The root directory for tables and charts is stored in the scripts.cfg file, variable TABDIR. This file also holds general configuration options used by the scripts. The plot-graph.sh script relies on the gnuplot script (flag -g) to plot the graph. Currently, the ones provided (extension .gnu) only support 8 threads, but it should not be hard to change them to your needs.

Please keep in mind that this is working-in-progress and any help is really welcome. Also, take a look at our artifact page for further information.

Known Issues

• When compiling the version of gperftools provided, you might run into a configuration issue due to a different version of automake. You can try to download and perform a local installation of automake 1.13 to solve the issue.

• Only if you need PAPI and malloc count support: You need to manually set up the BASEDIR variable in the tm-study-malloc/benchmarks/stamp/trunk/common/Makefile.common file with the absolute path of your base directory.