To reproduce our evaluation of the mitigation that switches LAM/UAI off/on upon kernel entry/exit (Table 4 of the paper), do the following.
We assume you are running an Ubuntu 22.04 installation on an Intel or AMD machine.
Download the Linux kernel 6.3 source code.
wget https://github.com/torvalds/linux/archive/refs/tags/v6.3.zip
unzip v6.3.zip
rm v6.3.zip
cd linux-6.3
Configure the kernel as we did (Ubuntu defaults). On Intel that is
cp ../switch-lam.config .config
On AMD that is
cp ../switch-uai.config .config
Build the kernel as a Debian package.
make -j`nproc` bindeb-pkg
Apply the switching patch. On Intel machines that is
git apply ../switch-lam.patch
while on AMD machines that is
git apply ../switch-uai.patch
Set the kernel's name.
sed -i 's/\(CONFIG_LOCALVERSION="\)/\1-switch/g' .config
Build this kernel as well.
make -j`nproc` bindeb-pkg
Now install both kernels.
dpkg -i ../linux-*.deb
Reboot into the kernel you want to benchmark first.
grub-reboot [kernel-number]
reboot
Download the LMBench source code.
git clone https://github.com/intel/lmbench.git
Fix this broken version of LMBench on Ubuntu.
git apply ../lmbench-ubuntu.patch
Configure, build, and run LMBench. We configured using the defaults, except for
setting the memory to use to 512MB, cf. lmbench-*.config for our config files.
make results
To run multiple rounds, use
make rerun
Next, switch to the other kernel (baseline vs switching) and repeat.
You can find our results in the results folder. The *.out files
contain the raw output of our LMBench runs. The summarize.py script dumps
Table 4 of the paper (Latex formatted) together with geomeans.