PTEMagnet Artifact Evaluation Pack

This repository includes an artifact evaluation pack for PTEMagnet (#111 ASPLOS'21 paper). The artifact consists of:

• a patch implementing PTEMagnet in Linux Kernel v4.19 together with scripts for building clean and modified versions of the Linux kernel
• a disk image for QEMU containing Ubuntu 16.04 LTS with selected SPEC'17, GPOP, and MLPerf ObjDetect benchmarks with their reference datasets; the disk image also contains scripts for running SPEC'17 and GPOP benchmarks colocated with MLPerf ObjDetect
• a python scripts for launching and measuring the execution time of SPEC'17 and GPOP benchmarks when running colocated with MLPerf ObjDetect within a virtual machine with and without PTEMagnet
• scripts for installing relevant tools and setting the environment

Installation

This code is designed to run on a two-socket x86 server with 20+ physical cores, more than 128GB of memory, and running Ubuntu 18.04 LTS.

Part 1: clone this repo

sudo apt-get update
sudo apt-get install git
git clone --recurse-submodules https://github.com/amargaritov/PTEMagnet_AE.git

Part 2: install packages and set environment

This step can take up to 2 hours.

cd PTEMagnet_AE
./install/install_all.sh <PATH_TO_DIR_WITH_AT_LEAST_150GB_FREE_SPACE>
source source.sh

This script

• installs all relevant tools & libraries
• builds the clean kernel and the modified kernel with PTEMagnet
• downloads the disk image for a VM (if run outside of CloudLab)
• sets relevant shell and python environment for scripts automating launching and measuring the execution time of benchmarks
• disables THP on the host machine

Note that we can provide access to a preconfigured CloudLab profile (and servers) on which this code was tested. CloudLab is a flexible, scientific infrastructure which provides researchers with control and visibility all the way down to the bare metal. Using the CloudLab profile accelerates installation and simplifies troubleshooting. If you are interested in running the artifact evaluation on CloudLab, please email Artemiy [email protected].

Part 3a: setting ssh keys for passwordless ssh

Evaluation scripts should be able to passwordlessly ssh to 1) a virtual machine where the benchmarks would run and 2) to the host. In our opinion, the simplest way to achieve passwordless ssh is using ssh keys. To upload an ssh key to the virtual machine

• Generate ssh key with

ssh-keygen -t rsa

• Boot a virtual machine with the provided disk image

KERNEL=clean; sudo qemu-system-x86_64 -kernel $KERNEL_DIR/linux_$KERNEL/arch/x86/boot/bzImage -boot c -m 64G -hda $IMAGE_DIR/rootfs.img -append "root=/dev/sda rw" -device e1000,netdev=net0 -netdev user,id=net0,hostfwd=tcp::6666-:22 --enable-kvm -smp 20 -cpu host -nographic

• Wait for machine to boot (login prompt should appear)
• Start a new shell on the host, upload the generated ssh key to a virtual machine

ssh-copy-id -p 6666 user@localhost

when asked for password, type user

• To allow passwordless login from the host to the host itself upload the ssh key to it

# This does not work on Cloudlab!
ssh-copy-id $USER@localhost

If using CloudLab, you would need to copy your cloudlab ssh key to the host

# on machine you ssh to Cloudlab from
scp -P 22 ~/.ssh/<YOUR_CLOUDLAB_KEY> <CLOUDLAB_USERNAME@CLOUDLAB_MACHINE>:~/.ssh/ 
# on Cloudlab machine
chmod 400 ~/.ssh/<YOUR_COPIED_CLOUDLAB_KEY>
eval "$(ssh-agent -s)"
ssh-add ~/.ssh/<YOUR_COPIED_CLOUDLAB_KEY>

Test that you can now passwordlessly login to the host

ssh $USER@localhost

• Shutdown the virtual machine

ssh -p 6666 user@localhost 'sudo shutdown -h now'

Part 3b: set up passwordless sudo on the host

• Open the /etc/sudoers file (as root, of course!) by running:

sudo visudo

• At the end of the /etc/sudoers file add this line: <username> ALL=(ALL) NOPASSWD:ALL

Part 4: Disable dynamic CPU power management

Note that this step is optional. This step should only be done if the measurements do not reproduce the expected results. For the initial setup, you can proceed to the Evaluation section.

We found that dynamic CPU frequency scaling migth lead to variability in performance and can introduce substational noise to the measurements. As a result, to reduce the system jitter in the latency measurement experiments it is better to disable frequency scaling (set constant CPU frequency). For fixing CPU frequency, one needs

• change BIOS settings
  • disable P states
  • setting "System Profile Settings" in BIOS -> “CPU power management” needs to be set to “ OS DBPM” instead of “Maximum performance” (the actual list of BIOS settings to change depends on a server model, please contact Artemiy [email protected] if you face problems -- we may provide a preconfigured Cloudlab machine to you)
• update Linux kernel boot paramenters to change the frequency driver from intel_pstate to acpi driver
  • edit /etc/default/grub: add the following boot parameters to GRUB_CMDLINE_LINUX_DEFAULT: usbcore.autosuspend=-1 intel_pstate=disable intel_iommu=on iommu=pt nokaslr rhgb quiet tsc=reliable cpuidle.off=1 idle=poll intel_idle.max_cstate=0 processor.max_cstate=0 pcie_aspm=off processor.ignore_ppc=1 As a result, you should have a line like this one `GRUB_CMDLINE_LINUX_DEFAULT="quiet splash usbcore.autosuspend=-1 intel_pstate=disable intel_iommu=on iommu=pt nokaslr rhgb quiet tsc=reliable cpuidle.off=1 idle=poll intel_idle.max_cstate=0 processor.max_cstate=0 pcie_aspm=off processor.ignore_ppc=1"
  • update grub settings

  sudo update-grub

  • reboot to the updated kernel

  sudo reboot
  

• install cpufreq

sudo apt-get install cpufrequtils

• check if intel_pstate got changed to acpi driver:

cpufreq-info | grep driver

should show acpi

• set frequency on all cores to 2GHz

$REPO_ROOT/install/freq/cpufreq-set-all -f 2000000

Evaluation

Getting all the results for Figure 6

You can launch the evaluation for all the benchmarks with launch_all_exps. launch_all_exps.sh will perform 25 execution time measurements for all 8 benchmarks on with 2 kernel versions. Firstly, create a screen session

screen -L -Logfile ae_all_log -S ae_ptemagnet_all

Then start the evaluation of all benchmarks

cd $REPO_ROOT/evaluation;
. launch_all_exps.sh

and make sure the script is running (the virtual machine should start booting, then after about 40 seconds it should move to prefaulting host memory, then launch MLPerf and a benchmark). If you get an import problem with distbenchr inside the screen run

cd distbenchr
python setup.py install
cd ..

and try running launch_all_exps.sh again.

Something got wrong if instead of virtual machine booting you get a prompt for logging in to the host machine (as the current host user) and the launch_all_exps.sh script doesn't output anything new to the console for more than 2 minutes after it was started: this shouldn't happen. If this is the case, please repeat setting the passwordless ssh for the host (Part #3 in Installation).

You can minimize (detach) the screen by pressing CTRL+A and D. You can now log out from the host machine as execution will take time (approximately 60 hours). We recommend checking how it is doing from time to time though. You can open a screen to see what is going on with

screen -r ae_ptemagnet_all

View the results

You can view the results generated by launch_all_exps.sh by running

cd $REPO_ROOT/evaluation
. show_results.sh

Running individual benchmarks (if needed)

cd ./evaluation/; 
mkdir -p results;
./lauch_exp.py --experiment_tag asplos21_ae --kernel [ clean | modified] --app [ bfs | cc | nibble | pr | gcc | mcf | omnetpp | xz ] --num_experiments <int>  --result_dir <FULL_PATH_DIR_TO_STORE_RESULT_FILES_TO> 

This script runs the benchmark specified under the app parameter in colocation with MLPerf ObjDetect in a virtual machine under the selected kernel type num_experiments number of times. The script outputs (prints) the average execution time of the benchmark in such an environment at the end of its execution. For example, the following command will print average execution time for mcf run 10 times on a kernel with PTEMagnet (in colocaiton with MLPerf ObjDetect):

cd $REPO_ROOT/evaluation/; mkdir -p results; ./launch_exp.py --experiment_tag asplos21_ae --kernel modified --app mcf --num_experiments 10 --result_dir $(pwd)/results

To reproduce the results of Figure 6, one needs to run the script for each benchmark two times: with clean and modified kernel and compare the execution times (launch_all_exps.sh automates this process). The launch_all_exps.sh script lauches each benchmark 25 times and evaluates two configurations, so its execution takes approximately 60 hours in total.

Miscellaneous

Building Linux kernel

Note that both clean and modified Linux kernels were already built by the installation script. You need to run it only if you want to rebuild them.

./install/build_kernels.sh <PATH_TO_WHERE_PUT_KERNELS> 

Note that both clean and modified Linux kernels were already built by the installation script. You need to run it only if you want to rebuild them.

Downloading the VM image

Note that the VM image is already downloaded by the installation script. You don't need to run it.

./install/download_vm_disk_image.sh <PATH_TO_WHERE_PUT_IMAGE>

Running a virtual machine manually

If you want to boot a virtual machine with the VM image manually you can run the following command:

KERNEL=clean; sudo qemu-system-x86_64 -kernel $KERNEL_DIR/linux_$KERNEL/arch/x86/boot/bzImage -boot c -m 64G -hda $IMAGE_DIR/rootfs.img -append "root=/dev/sda rw" -device e1000,netdev=net0 -netdev user,id=net0,hostfwd=tcp::6666-:22 --enable-kvm -smp 20 -cpu host -nographic

This command boot a machine with 20 cores and 64GB memory. After launching the virtual machine you can ssh into it with

ssh -p 6666 user@localhost

password is user