starduster / gem5-bc-workload

title	author
gem5 quickstart	Jason Lowe-Power

Running the simulation

If you have docker on your system, you can use the following command.

docker run --rm -u $UID:$GID -v `pwd`:`pwd` -w `pwd` gcr.io/gem5-test/gem5-all-min-dependencies run.py

To use singularity instead, you can use the following command.

singularity run docker://gcr.io/gem5-test/gem5-all-min-dependencies run.py

The simulator will output the simulated time, the total instructions, and the total cycles. For more detailed statistics, you will find m5out/stats.txt contains many detailed statistics.

By default, this runs a matrix multiple with ROI annotations and gives the number of cycles, instructions, and simulated time for only the ROI.

You don't need to modify anything to get started. If you want to modify the workload implementation or modify the simulator, you can see below for details on how everything fits together.

Options for the runscript

You can find details on the options you can pass to this simulation script by using --help.

docker run --rm -u $UID:$GID -v `pwd`:`pwd` -w `pwd` gcr.io/gem5-test/gem5-all-min-dependencies run.py --help

Choosing an output directory

When running the simulator and performing many different experiments, you may want to use different directories for the output. To do this, you can set the "outdir" on the command line. Note that this is a gem5 parameter, not a parameter to the script, so you need to set the parameter before run.py on the command line. See below for an example.

docker run --rm -u $UID:$GID -v `pwd`:`pwd` -w `pwd` gcr.io/gem5-test/gem5-all-min-dependencies --outdir=ooo-stats/ run.py --processor_type=out-of-order

Workload

The workload is a simple matrix multiply as shown below.

for (int i = 0; i < size; i++) {
    for (int k = 0; k < size; k++) {
        for (int j = 0; j < size; j++) {
            C[i][j] += A[i][k] * B[k][j];
        }
    }
}

You can find the code in the mm.cpp file.

Two binaries are included in this directory:

• mm: An x86 binary which can be used for testing
• mm-gem5: An x86 binary with gem5 "magic" instructions which delineate the region of interest. You should use this binary with gem5.

Using the workload in gem5

The file workload.py defines a gem5 workload which loads the mm-gem5 binary into the default SimpleX86Board and the OutOfOrderProcX86Board. You can change the input size by modifying the workload.py file.

By default when you run with the run.py script, some main statistics are printed on the terminal output. If you want to see all of the detailed statistics you can find them in the m5out/stats.txt file.

Building the workload

To compile the binary for your native machine, you can use make mm.

If you want to build the binary with the region of interest markers for gem5, you must first make sure that you have the gem5 source downloaded locally. You will also have to ensure that the gem5 m5 library is built. To build the m5 library, you can use the following command.

cd gem5/util/m5
scons build/x86/out/libm5.a

Then, to build mm-gem5 you can use make mm-gem5. You may need to pass the base gem5 directory with make mm-gem5 GEM5_ROOT="<path to gem5>"

Downloading/building gem5

You can download gem5 from the gem5 git repository as shown below.

git clone https://gem5.googlesource.com/public/gem5

To build gem5 you will need to install all dependencies and then build gem5. See the gem5 website for details. Once you have the dependencies installed, you can use the following command.

git checkout develop
scons -j`nproc` build/ALL/gem5.opt

Note that this command checks out the develop branch, which is the most up to date version of gem5. Until gem5-v22.1 is released, this repository will only work with the develop branch. After gem5-v22.1 is released, you can use the default, stable, branch.

Note that you only need to download gem5 if you want to build the mm-gem5 binary yourself or if you want to make source code modifications to gem5.