DAIET: A P4 repository from AmedeoSapio

DAIET: Data Aggregation In nETwork

Overview:

Many scalable data center applications follow a partition-aggregate pattern where data and computations are distributed among many servers and their partial results are exchanged over the network and aggregated to produce the final output. For these workloads, the network communication costs are often one of the dominant scalability bottlenecks.

To improve performance, we present DAIET, a system that performs data aggregation along network paths by leveraging emerging programmable network devices. By offloading part of the aggregation task to the networking hardware, DAIET not only reduces the traffic at each network device, but also allows to reduce the computation done by the CPU, effectively decreasing job completion times. We describe a preliminary prototype of DAIET using P4 to perform in-network aggregation for MapReduce-based applications. Experimental results show that our approach achieves around 88% data reduction ratio and a similar decrease in the reducer running time, without requiring severe application-level modifications.

If you use DAIET in your research, please cite our paper:

@inproceedings{sapio2017daiet,
title={{In-Network Computation is a Dumb Idea Whose Time Has Come}},
author={Amedeo Sapio, Ibrahim Abdelaziz, Abdulla Aldilaijan, Marco Canini, and Panos Kalnis},
booktitle={Proceedings of the 16th ACM Workshop on Hot Topics in Networks (HotNets)},
year={2017},
organization={ACM}
}

Resources

DAIET project website: sands.kaust.edu.sa/daiet/

ACM HotNets 2017 Experiment:

Note: tested with Ubuntu 17.04 Zesty Zapus (amd64) on May, 2017.

Install dependencies: net-tools, git, python, python-pip, screen, maven, openjdk-8-jdk, docker

sudo apt-get update
sudo apt-get install -y \
    linux-image-extra-$(uname -r) \
    linux-image-extra-virtual \
    git python python-pip screen maven net-tools \
    openjdk-8-jdk \
    apt-transport-https ca-certificates curl \
    software-properties-common

export JAVA_HOME=/usr/lib/jvm/java-8-openjdk-amd64/
export PATH=$PATH:/usr/lib/jvm/java-8-openjdk-amd64/bin

# Install docker
curl -fsSL https://download.docker.com/linux/ubuntu/gpg \
    | sudo apt-key add -

# Currently there is no docker repo for Zesty, 
# so we use Xenial’s.
sudo add-apt-repository \
   "deb [arch=amd64]
   https://download.docker.com/linux/ubuntu \
   xenial stable"

sudo apt-get update
sudo apt-get install -y docker-ce
sudo service docker start

Download the P4 compiler repository and install it:

wget https://github.com/p4lang/p4c-bm/archive/1.9.0.zip
unzip 1.9.0.zip
cd p4c-bm-1.9.0
sudo pip install -r requirements.txt
sudo pip install -r requirements_v1_1.txt
sudo python setup.py install
cd ..

Clone the DAIET and DAIET-MapReduce repositories:

git clone https://github.com/AmedeoSapio/DAIET.git
git clone https://github.com/ibrahimabdelaziz/DAIET-MapReduce.git

Compile DAIET:

cd DAIET
p4c-bmv2 p4src/daiet.p4 --json daiet.json
cd ..

Compile DAIET-MapReduce:

cd DAIET-MapReduce 
mvn package
chmod +x dist/bin/*
cd ..

Extract the dataset:

unzip DAIET/tests/data/random_text_500mb.zip

We pin the container running the switch on the cores {4,5,6,7} and the 12 hosts containers on the cores from 8 to 31. You must change the parameter: “--cpuset-cpus” twice in “DAIET/tests/single_switch/Experiment.py” if you don’t have enough cores. We also isolated those cores from kernel scheduling using the “isolcpus” kernel parameter.
Optional: To run the experiment using a RAMdisk, we changed Docker's storage base directory (where container and images go) using the -g option when starting the Docker daemon. You can mount a RAMdisk (tmpfs) on the folder “/memfs” and then edit the “/etc/default/docker” file with the -g option:

# In /etc/default/docker
DOCKER_OPTS="-dns 8.8.8.8 -dns 8.8.4.4 -g /memfs"

Restart docker afterwards:

service docker restart

Start the testbed:

wd=`pwd`
cd DAIET/tests/single_switch/
./Experiment.py ${wd}

Run a job once without using DAIET. This is needed because DAIET-MapReduce assigns the reduce tasks (and thus the reducer ID used as tree ID in the packets) randomly among the workers. We can see the assignment in a first job and predict the assignment for the next job, which follows the same order. In a future version, the jobtracker would push the rules in the switch when a new reduce task is scheduled.

sudo screen -r h13
# Now you are in the master container

# Load the data file in the DFS
./DAIET-MapReduce/dist/bin/dfs-load -rh h13 -rp 7777 -r 12 \
    random_text_500mb.txt

# Start a job without DAIET (24 mappers, 12 reducers)
./DAIET-MapReduce/dist/bin/examples-wordcount \
    random_text_500mb.txt result.txt -m 24 -r 12 -rh h13 \
    -rp 7777

# Check the job status
./DAIET-MapReduce/dist/bin/mapreduce-jobs -rh h13 -rp 7777

Wait for the end of the job and then press “CTRL+A” and “D” to exit from the master container.

Extract the reducer IDs from the logs:

sudo ./extract_reducer_id.sh

This command prints 12 integers. You have to add 36 (number of mapper and reducers) to each one of them and use the result to update the file “DAIET/tests/single_switch/commands.txt” (replace “XX“ with the new values). Specifically, the integers must be used as matching value in the “tree_id_adapter_table” and “forwarding_table”.

Add the flow rules to the switch:

sudo screen -r bmv2
# Now you are in the switch container

simple_switch_CLI \
    < /Testbed/DAIET/tests/single_switch/commands.txt

Press “CTRL+A” and then “D” to exit from the switch container.

Run a job with DAIET. Note that the workers are connected to the switch through the interfaces from sw-h1 to sw-h12 (sw-h13 is the master, which is not a worker).

sudo screen -r h13
# Now you are in the master container

# Start a job with DAIET (24 mappers, 12 reducers)
./DAIET-MapReduce/dist/bin/examples-wordcount \
    random_text_500mb.txt result.txt -m 24 -r 12 -rh h13 \
    -rp 7777 -nr

# Check the job status
./DAIET-MapReduce/dist/bin/mapreduce-jobs -rh h13 -rp 7777

Wait for the end of the job and then press “CTRL+A” and “D” to exit from the master container.

The time spent for the reduce phase (in milliseconds) can be printed with:

sudo ./extract_reduce_time.sh

It prints, for each worker, the reducer ID and the reduce time for both jobs, with and without DAIET. The first job have reducer IDs in [24,35] and the second in [60,71].

Finally, clean the testbed:

./Experiment.py clean