Mitosis ASPLOS'20 Artifact Evaluation

This repository contains scripts for the ASPLOS'20 artifact evaluation of the paper Mitosis: Transparently Self-Replicating Page-Tables for Large-Memory Machines by Reto Achermann, Jayneel Gandhi, Timothy Roscoe, Abhishek Bhattacharjee, and Ashish Panwar.

The scripts in this paper can be used to reproduce the figures in the paper.

Authors

• Reto Achermann (ETH Zurich)
• Jayneel Gandhi (VMWware)
• Timothy Roscoe (ETH Zurich)
• Abhishek Bhattacharjee (Yale University)
• Ashish Panwar (IISc Bangalore)

License

See LICENSE file.

Directory Structure

• precompiled contains the downloaded binaries
• build contains the locally compiled binaries
• sources contains the source code of the binaries
• datasets contains the datasets required for the binaries
• scripts contains scripts to run the experiments
• bin points to the used binaries for the evaluation (you can use scripts/toggle_build.sh to switch between precompiled and locally compined binaries)

Hardware Dependencies

Some of the workingset sizes of the workloads are hardcoded in the binaries. To run them, you need to have a multi-socket machine with at least 128GB of memory per NUMA node. e.g. 4 socket Intel Xeon E7-4850v3 with 14 cores and 128GB memory per-socket (512 GB total memory)

Software Dependencies

The scripts, compilation and binaries are tested on Ubuntu 18.04 LTS. Other Linux distributions may work, but are not tested.

In addition to the packages shipped with Ubuntu 18.04 LTS the following packets are required:

sudo apt-get install build-essential libncurses-dev \
                     bison flex libssl-dev libelf-dev \
                     libnuma-dev python3 python3 python3-pip \
                     python3-matplotlib python3-numpy \
                     git wget kernel-package fakeroot ccache \
                     libncurses5-dev wget pandoc libevent-dev \
                     libreadline-dev python3-setuptools

In addition the following python libraries, installed with pip

pip3 install wheel
pip3 install zenodo-get

Docker There is a docker image which you can use to compile. You can do make docker-shell to obtain a shell in the docker container, or just to compile everything type make docker-compile.

Obtaining Pre-Compiled Binaries

This repository does not contain any source code or binaries. There are scripts which download the pre-compiled binaries, or source code for compilation.

Obtaining Pre-Compiled Binaries

To obtain the pre-compiled binaries execute:

./scripts/download_binaries.sh

The pre-compiled binaries are available on Zenodo.org. You can download them manually and place them in the precompiled directory.

There are several binaries available on Zenodo:

• bench_* are the benchmarks used in the paper
• page_table_dump/numactl/perf are helper utilities for collecting stats
• linux-*.deb are the linux kernel image and headers with Mitosis modifications.

Obtaining Source Code and Compile

If you don't want to compile from scratch, you can skip this section.

The source code for the Linux kernel and evaluated worloads are available on GitHub. To obtain the source code you can initialize the corresponding git submodules. Note: the repositories are private at this moment, as they are not ready for public release.

git submodule init
git submodule update

To compile everything just type make

To compile the different binaries individually, type:

• Mitosis Linux Kernel: make mitosis-linux
• Mitosis numactl: make mitosis-numactl
• BTree: make btree
• Canneal: make canneal
• Graph500: make graph500
• GUPS: make gups
• HashJoin: make hashjoin
• LibLinear: make liblinear
• PageRank: make pagerank
• Redis: make redis
• XSBench: make xsbench
• memops: make memops
• memcached: make memcached

Evaluation Preparation

To run the evaluations of the paper, you need a suitable machine (see Hardware Dependencies) and you need to boot your machine with the Mitosis-Linux you downloaded or compiled yourself. Both, the kernel image and the headers!.

To install the kernel module for page-table dumping you need to execute:

make install lkml

It's best to compile it on the machine runnig Mitosis-Linux.

make mitosis-page-table-dump

Deploying

Just clonse the artifact on the machine you want to run it on.

For deploying on a remote machine only.

To deploy the binaries and scripts on a remote machine, just clone the repository locally, and specify the target host-name you want to run the artifact on in ./scripts/site_config.sh. Then run the following script locally.

./scripts/deploy.sh

Preparing Datasets

Some workloads require datasets to run. Scripts to download or generate the datasets are placed in datasets/. These datasets require approximately 100GB of disk space. Generate datasets as:

datasets/prepare_liblinear_dataset.sh
datasets/prepare_canneal_datasets.sh

Running the Experiments

Before you start running the experiments, make sure you fill in the site configuration file site-config.sh.

To run all experiments, execute (this may take a while...)

scripts/run_all.sh

To run the experiments for a single figure, do:

• Figure 6 - ./scripts/run_f6_all.sh
• Figure 9a - ./scripts/run_f9a_all.sh
• Figure 9b - ./scripts/run_f9b_all.sh
• Figure 10a - ./scripts/run_10a_all.sh
• Figure 10b - ./scripts/run_10b_all.sh
• Figure 11 - ./scripts/run_f11.sh
• Table 5 - ./scripts/run_t5.sh

You can also execute each bar of Figure-6, Figure-9 and Figure-10 separately. For Figure-6 and Figure-10, execute as:

./scripts/run_f6f10_one.sh BENCHMARK CONFIG

For Figure-9, execute as:

./scripts/run_f9_one.sh BENCHMARK CONFIG`

Naming conventions for arguments:

• Use "small letters" for benchmark name (e.g., btree, xsbench).
• Use "CAPITAL LETTERS" for configuration name (e.g., TLPLD, RPILDM).

Refer to ./scripts/run_f6_all.sh and ./scripts/run_f10a_all.sh for more examples on how to execute a single benchmark configuration.

All output logs will be redirected to "evaluation/measured/$FIGURENUM".

To process the logs for Figure-6, Figure-9 and Figure-10, execute:

./scripts/process_logs_fig_6-9-10.py`

Collecting Experiment Data

In case you used the deploy script, you can execute

./scripts/collect-results.sh

To copy the results from the remote machine to your local one.

Compare the Experiments

When you collected all the experimental data, you can compare them against the reference data shown in the paper:

./scripts/compile_report.sh

Paper Citation

Reto Achermann, Ashish Panwar, Abhishek Bhattacharjee, Timothy Roscoe, and Jayneel Gandhi. 2020. Mitosis: Transparently Self- Replicating Page-Tables for Large-Memory Machines. In Proceed- ings of the Twenty-Fifth International Conference on Architectural Support for Programming Languages and Operating Systems (ASP- LOS ’20), March 16–20, 2020, Lausanne, Switzerland. ACM, New York, NY, USA, 18 pages. https://doi.org/10.1145/3373376.3378468