The Impact of Meta-Tracing on VM Design and Implementation

Carl Friedrich Bolz and Laurence Tratt

Repeatable experiments

Our paper contains experiments comparing 11 synthetic benchmarks over various VMs. We believe strongly in repeatability in computing experiments, and thus provide the same automated scripts we used for the experiment. Our system automatically builds the same version of the VMs from the paper (except, for licensing reasons, HotSpot) and runs the experiments. Our hope is that others can easily run our experiments and study the output on their particular setup.

Versions available are:

The full source code of the experiment is also available at our GitHub repo.

Results

We are interested to see how the experiment runs on different machines. If you have access to an idle machine with a few days of CPU time, we welcome your running the experiment and sending is your results.html file for inclusion here.

Building and running

We have attempted to make this as automated as possible on Unix machines. We recommend that the machine you run this on has at least 6Gb RAM (preferably 8Gb). You may need to install some background software (the build process detects some dependencies, but you may need to install various libraries e.g. pcre). The only piece of software we can not easily automate the download of is Java: you will need to provide your own Java VM (including javac) and put it into your $PATH.

In general you can simply execute make:

  make
This will build the necessary software (which takes approximately 1-2 hours) and then run the experiments. These default to 30 repetitions of each benchmark, for a rough running time of 3-7 days. If you want a shorter run, you can edit the REPS parameter at the top of run_benchmarks.sh (minimum value is 1). When complete, you will see the raw output from multitime in the results file; if you have Python installed, a pretty-printed HTML version of the results can be found in results.html.

Known issues

Although our experiment attempts to fully automate the building and running of our experiment, portability is always an issue.

LoC

We have also automated the Lines of Code (LoC) count for VMs so that you can see precisely what we have included and excluded. In general you can simply execute:
  $ ./loc.sh
Note that if "build.sh" has not previously been run, "build.sh" will be executed to download and build all the VMs (though the experiments themselves will not be run).