biswapanda / timeloop

Timeloop user guide

Getting started

• Install the following dependencies.

scons
libconfig++-dev
libboost-dev
libboost-iostreams-dev
libboost-serialization-dev
libyaml-cpp-dev
libncurses-dev
libtinfo-dev
libgpm-dev

• Clone the timeloop repository.

mkdir timeloop-dev
cd timeloop-dev
git clone ssh://path/to/timeloop.git

• In addition to the main source code, you need the source code for a power-area-timing (pat) model to build timeloop. A placeholder pat model is included in the main repository. Before building timeloop, place a symbolic link to the pat model like so:

cd timeloop/src
ln -s ../pat-public/src/pat .
cd ..

• Instead of the included placeholder pat model, you may build against any other custom pat model, as long as it exports the same interface as the pat/pat.hpp in the included model. The implementation must be in a pat/pat.cpp file. As before, create a symbolic link to the source code for the power-area-timing model and place it in src/pat, for example:

git clone ssh://path/to/timeloop-pat[XXX].git
cd timeloop/src
ln -s ../../timeloop-pat[XXX]/src/pat .
cd ..

• Another way to provide a power/energy model to Timeloop is to integrate Accelergy with Timeloop. To do so, you need to either install Accelergy so that the shell can find it (i.e., which accelergy works), or provide the path to Accelergy binary as an environmental variable, ACCELERGYPATH, when running Timeloop.

When building timeloop in the next step, you also need to provide an extra --accelergy flag to scons so that it builds Timeloop and makes it use Accelergy for the energy model.

scons --accelergy

• Once the pat link is set up, you can build timeloop using scons.

scons -j4

This builds 3 different tools:

• timeloop-mapper is the complete application that instantiates an architecture, constructs its mapspace, searches for an optimal mapping within the mapspace and reports statistics for the optimal mapping.

• timeloop-model instantiates an architecture, evalutes a specific given mapping of a workload and reports the statistics.

• timeloop-metrics simply instantiates an architecture and reports its workload-independent characteristics such as area and energy-per-access for various architectural structures.

• By default, the scons script will use shared (dynamic) linking. The timeloop libraries will be placed in the lib/ subdirectory. You can manually add that to LD_LIBRARY_PATH, or if you are using bash you can just source the provided environment setup script:

source env/setup-env.bash

• Run timeloop with a sample configuration.

cd configs/mapper
../../build/timeloop-mapper ./sample.yaml > sample.out

This will place timeloop's log in sample.out and generate the following outputs:

• timeloop-mapper.stats.txt Simulation stats (performance, energy, etc.)
• timeloop-mapper.map.txt/cfg The optimal mapping in different formats (the latter can be used in conjunction with the input architecture and problem spec to re-run the model on the optimal mapping.)
• timeloop-mapper.map+stats.xml An XML-formatted copy of the stats and optimal mapping which is used by various Python scripts to extract results from batch runs.

Further reading

Serially walking through the exercises in our Timeloop tutorial series serves as an excellent hands-on introduction to the tool.

For a deeper technical overview please read our ISPASS 2019 paper.