Skip to content
This repository has been archived by the owner on Nov 9, 2022. It is now read-only.

Latest commit

 

History

History
234 lines (169 loc) · 6.93 KB

README.md

File metadata and controls

234 lines (169 loc) · 6.93 KB
Raw

Experiment Testbed 🛌

This directory contains all the necessary code to run experiments using the blocklearning library via Docker.

Setup

Generate Ethereum Accounts

Generate the accounts that will be used on the network. To generate 5 accounts for miners (or validators) and 10 for trainers, run:

python3 toolkit.py generate-accounts 10 25

The files in ethereum/datadir are as follows:

  • keystore includes the accounts information that were generated and this is necessary to start the node.
  • geth/nodekey_owner is the node key that will be given to the "owner" RPC endpoint for our Ethereum network.
  • geth/nodekey_{i} is the node key that will be given to the i-th Ethereum miner.
  • geth/static-nodes.json includes the addresses that will be added as static peers to our nodes.
  • accounts.json is the account information and password.
  • miners.json has the public addresses generated from the private key of the miners nodekey. This is important for some consensus protocol genesis.

Update Genesis with Accounts

After generating the accounts, the genesis files needs to be generated with the new accounts accounts in order to boot the network with 100 ETH in each of the accounts. To do so, run:

python3 toolkit.py update-genesis

Build Docker Images

Several Docker images are used to run the Ethereum blockchain. To build then, run:

python3 toolkit.py build-images

Create Docker Network

docker network create \
  --driver=bridge \
  --subnet=172.16.254.0/20 \
  bflnet

Launch Blockchain Containers

For Docker compose, use:

CONSENSUS=poa MINERS=10 docker compose -f blockchain.yml -p bfl up

Where CONSENSUS = poa|qbft|pow.

Connect Peers

Unfortunately, peer discovery doesn't work with private networks. Not even if we use a bootstrap node. Thus, we need to connect the peers to each other manually.

python3 toolkit.py connect-peers <network>

Where <network> is the ID of the Docker network where the containers are running. You can check that by running docker network ls and looking for priv-eth-net. If no network is passed, the script will try to infer the correct network.

Deploy the Contract

The contract deployment script fetches the account to use from ethereum/datadir/accounts.json and uses the first account (index 0).

python3 toolkit.py deploy-contract

Launch ML Containers

CONTRACT=0x8C3CBC8C31e5171C19d8a26af55E0db284Ae9b4B \
  DATASET=mnist MINERS=10 AGGREGATORS=10 SCORERS=0 TRAINERS=25 \
  SCORING="none" ABI=NoScore \
  docker compose -f ml.yml -p bfl-ml up

Collect Statistics

Start collecting statistics before running the rounds (on the results repository):

DIR=./results/CURRENT/stats
mkdir -p $DIR

while true; do
  echo "fetching"
  docker stats --no-stream --format "{{.ID}}, {{.Name}}, {{.CPUPerc}}, {{.MemUsage}}, {{.MemPerc}}, {{.NetIO}}, {{.BlockIO}}" > $DIR/$(date '+%s').log
  sleep 2
done

Run Rounds

python3 start_round.py \
  --contract 0x8C3CBC8C31e5171C19d8a26af55E0db284Ae9b4B \
  --abi ../build/contracts/NoScore.json \
  --rounds 50

Collect Logs

To collect the logs of the trainers and validators afterwards, run:

python3 toolkit.py collect-logs

Contract

Some required base information for the contract can be found in ../contracts.json. This file includes two fields that must be filled before deploying the contract:

  • model: the IPFS CID of the exported model (head model for Split-CNN contract) in .h5 format.
  • bottomModel: the IPFS CID of the exported bottom model for the Split-CNN contract in .h5 format.
  • weights (optional): the IPFS CID with the initial weights.

IPFS

To add any file to IPFS, run:

ipfs add [-r] path

How to Run Different Experiments

Consensus Algorithms

CONSENSUS=poa|pow|qbft MINERS=10 docker compose -f blockchain.yml -p bfl up

CONTRACT=0x8C3CBC8C31e5171C19d8a26af55E0db284Ae9b4B \
  DATASET=mnist MINERS=10 SERVERS=10 CLIENTS=25 \
  SCORING="none" ABI=NoScore \
  docker compose -f ml.yml -p bfl-ml up

python3 start_round.py \
  --contract 0x8C3CBC8C31e5171C19d8a26af55E0db284Ae9b4B \
  --abi ../build/contracts/NoScore.json \
  --rounds 50

Selection Mechanisms

CONSENSUS=poa MINERS=10 docker compose -f blockchain.yml -p bfl up

CONTRACT=0xA63052D2C43CD996731937AAD7986bF8f88C2511 \
  DATASET=mnist MINERS=10 SERVERS=10 CLIENTS=25 \
  SCORING="none" ABI=FirstComeFirstServed \
  docker compose -f ml.yml -p bfl-ml up

python3 start_round.py \
  --contract 0xA63052D2C43CD996731937AAD7986bF8f88C2511 \
  --trainers fcfs \
  --abi ../build/contracts/FirstComeFirstServed.json \
  --rounds 50

Without Score Mechanisms

CONSENSUS=poa MINERS=10 docker compose -f blockchain.yml -p bfl up

CONTRACT=0x8C3CBC8C31e5171C19d8a26af55E0db284Ae9b4B \
  DATASET=mnist MINERS=10 SERVERS=10 CLIENTS=5|10|25|50 \
  SCORING="none" ABI=NoScore \
  docker compose -f ml.yml -p bfl-ml up

python3 start_round.py \
  --contract 0x8C3CBC8C31e5171C19d8a26af55E0db284Ae9b4B \
  --abi ../build/contracts/NoScore.json \
  --rounds 50

With Score Mechanisms

CONSENSUS=poa MINERS=10 docker compose -f blockchain.yml -p bfl up

CONTRACT=0x2988207C0b2666E554A803D25524B0822bd1B1A8 \
  DATASET=mnist MINERS=10 SERVERS=10 CLIENTS=5|10|25|50 \
  SCORING="<mechanism>" ABI=Score \
  docker compose -f ml.yml -p bfl-ml up

python3 start_round.py \
  --contract 0x2988207C0b2666E554A803D25524B0822bd1B1A8 \
  --abi ../build/contracts/Score.json \
  --scoring "<mechanism>" \
  --rounds 50

Vertical (Split-CNN)

CONSENSUS=poa MINERS=10 docker compose -f blockchain.yml -p bfl up

CONTRACT=0x4A59e668c68c7915bCdfD5530B7C1C3D0F83885f \
  DATASET=../../blocklearning-results/datasets/vertical-mnist MINERS=10 SERVERS=5 CLIENTS=2 \
  docker compose -f vml.yml -p bfl-ml up

python3 start_vertical_round.py \
  --contract 0x4A59e668c68c7915bCdfD5530B7C1C3D0F83885f \
  --rounds 50