dist-ml-orch

Implementation of distributed orchestration for federated learning. This work was done as part of master's thesis Distributed machine learning function orchestration in 5G networks.

This implementation has 3 components:

Containernet v3.1 is used to create a topology with Federated clients and servers as end hosts

Distributed MLFO is extension of Machine learning function orchestrator based on framework specified in ITU-T Y.3172

abh15/flower allows to start federated clients and servers with specified configuration via a REST interface

Prerequisites

Ubuntu 20.04

128 core CPU

Clone Containernet v3.1 and install using Option 1 Bare-metal installation

Install ONOS v2.6.0

Start ONOS

Terminal 1

cd onos

bazel run onos-local -- clean debug

Terminal 2

cd onos

tools/test/bin/onos localhost

onos> app activate org.onosproject.openflow

onos> app activate org.onosproject.fwd

Usage

Create topology

1. Copy scripts directory to containernet directory

2. Make sure promconfig in topo.py is set to correct path

3. cd containernet/scripts

   In terminal 1 excute the following and wait till the topology is created

   sudo python3 topo.py <num of cohorts> #e.g sudo python3 topo.py 5

   In terminal 2 execute following to update cpu limits and start background traffic application

   bash init.sh <num of cohorts> #e.g bash init.sh 5

   In terminal 3 execute following to test ping reachibility of topology nodes. Note that this may take long time. You may skip this step

   bash pingtest.sh

Send intent

We send intents to MLFO from a remote machine. Execute the following to send intents to MLFO and to start the FL process

cd intent

python3 sendintent.py

Check which cohort configuration are you sending. Details of cohort configuration are as follows

mlfoadd gives address:mlfo exposed port of remote machine where the mlfo is running

cohortdistr specifies how 50 clients are divided into cohorts (e.g 30,10,10 or 10,10,10,10,10). Number of cohorts should be same as that created in topology.

intentdist specifies which intent yaml to send for which cohort. Currently we only have 3 intents which work

sameserverdist specifies if multiple cohorts should use same FL server. Currently this is ignored

avgalgodist specifies which aggregation algorithm to use. Currently only FedAvg is supported

fracfit fraction of clients to use for fitting

minfit Minimum num of clients to fit

minav Minimum number of clients required to start FL rounds

numround Number of FL rounds

Send meter

This script sets meter for FL clients uplink at 100Mbps and downlink at 100Mbps. IN_PORT number in line 25 needs to be set to correct cohort number which needs to be metered(e.g 1, 2, 3)

sudo python3 sendmeter.py

Monitoring

1. Add Grafana.

   docker run -d -p 3000:3000 grafana/grafana

2. Navigate to http://10.66.2.142:3000. Go to configuration > add data source > add prometheus. Add the following HTTP URL and click save and test.

   http://10.66.2.142:9090

3. Add the dashboard by importing monitoring/dashboard.json file.

4. Open dashboard (das): http://10.66.2.142:3000

Note that prometheus.yml consists of configration which decides which statistics will be pulled from which containers. You may need to comment/uncomment depending on the requirement.

Added flows can be checked from the ONOS UI. http://10.66.2.142:8181/onos/ui/#/device

Misc commands

Reset mininet

sudo mn -c

Port mapping

Flower:

	5000 —> REST

    6000 —> FLWRINTERNAL

MLFO:

	8000 —> REST

  	9000 —> MLFO SERVER

Send test intent

curl -v -F file=@intent.yaml 'http://localhost:8000/receive'

Open bash shell at one of the nodes

sudo docker exec -it mn.fed.1 /bin/bash

sudo docker exec -it mn.fc.11 /bin/bash

sudo docker exec -it mn.mo.1 /bin/bash