salvacorts / gocey

🍀 P2P Distributed Evolutionary Algorithms on Ephemeral Infrastructure for Neural Network Optimization.

Bachelor Thesis: Gocey. Distributed Evolutionary Algorithms on Ephemeral Infrastructure

Author: Salvador Corts Sánchez

Supervisor: Juan Julián Merelo Guervós

Thesis Document: https://github.com/salvacorts/TFG-Thesis

---

Requirements

• Go >= 1.12: https://golang.org/doc/install
• Go Protocol Buffers: https://github.com/golang/protobuf

Go dependecies

go get github.com/salvacorts/TFG-Parasitic-Metaheuristics
cd $GOPATH/src/github.com/salvacorts/TFG-Parasitic-Metaheuristics/mlp-ea-decentralized/native
go get ./...

Run

All the following paths are relative to $GOPATH/src/github.com/salvacorts/TFG-Parasitic-Metaheuristics/mlp-ea-decentralized directory.

Island

cd native/server
go run -grpcPort <port> -clusterPort <port> -metricsPort <port> -clusterBoostrap <boostrap node addr:port> -datasetPath <path to dataset> -webPath <path to web> -webPort <port>
# e.g. go run server.go -grpcPort 2006 -clusterPort 4003 -metricsPort 5003 -clusterBoostrap 127.0.0.1:4001 -datasetPath ../../../datasets/glass.csv -webpath ../../web/src/ -webPort 8080

Evaluator

Native

cd native/client
go run client.go -server <island addr:port>
# e.g. go run client.go -server 127.0.0.1:2006

Browser (WebAssembly)

Open a new tab in yout browser and enter an island IP and its webPort port. E.g. if an island is running at 127.0.0.1, and its webPort parameter is 8080, go to 127.0.0.1:8080.

Modifying Go Browser Client Source: You will need to rebuild the webassembly file

cd web/src/go
GOOS=js GOARCH=wasm go build -o wasm/main.wasm

gocey's Issues

Distributed EA

Design and implement the communication schema between the private network (nodes in the cloud) and the rest of users (using their browsers)

As a possible contributor, I want to have all functions tested so that my own contributions can be checked before being added.

User Story

Tests are essential, and they must be included along with the code, at the same time if possible.

Acceptance criteria

If possible, reach 90% coverage in tests.

Definition of Done

• Code Builds without error warning
• Code is Unit Tested
• Code is Deployed to system test environment and passed system tests
• Build pushed to a demo server
• Documentation

As an Administrator, I want to run a demo problem so that I can test the platform.

User Story
I need a problem that can potentially perform well in the platform.
I will use this problem to test the platform while I develop it.

Acceptance criteria

• Fitness function
• Mutation function
• Crossover function
• Selection function

Definition of Done

• Code Builds without error warning
• Code is Unit Tested
• Code is Deployed to system test environment and passed system tests
• Build pushed to a demo server
• Documentation

Representación de soluciones

Dado que de problema a problema la representación de una solución puede ser muy diferente, ¿cómo podríamos transmitir soluciones independientemente de su representación entre un nodo y el servidor?

Metaheurísticas a implementar

Aunque idealmente el sistema debería abstraerse del problema a resolver, sería conveniente discutir que algortimos vamos a implementar a fin de planificarnos y saber que restricciones tendrá el sistema.

As a tech-savvy user, I want to contribute to the problem by running code natively so that I can achieve better performance.

User Story
I want to run code on a machine installing whatever is necessary but achieving better performance than from the browser.

I want to be able to communicate all my machines running code with other machines running code natively in a decentralized way so there is no single point of failure.

I also want my execution to be coordinated with the rest of the collaborators on the platform by letting them ask my machines for information related to the execution of the problem.

Acceptance criteria
Provide a middleware that provides:

• Epidemic protocol to share state across machines on the platform about nodes available to share information
• Share information about solutions with other nodes when requested
• Communicate in a way as decentralized as possible while still being able to work with browser-based contributors

Definition of Done

• Code Builds without error warning
• Code is Unit Tested
• Code is Deployed to system test environment and passed system tests
• Build pushed to a demo server
• Documentation

As an examiner, I want to know the motivations of the work so that I can evaluate it properly.

User Story

The examiners need to know several things

• Motivation for the system
• Examination of the state of the art
• Definition of clear, reachable and measurable objectives.

Acceptance criteria

Discussion of all issues involved in this issue or in person, and eventually writing them in the report.

Definition of Done

Introductory chapter to memory completed.

As a researcher, I want to upload problems so that users can collaborate to it.

User Story
As a researcher I want to be able to upload my problem to the platform so contributors can select it from the website.

Contributors need information about the problem they are contributing to; as a researcher, I want to provide information related to the problem to convince users to contribute to my problem as well as informing them about what is going on on their browsers.

Acceptance criteria

• Upload code that I want to run in the browser
• Fill a form about the problem that contains :
  • Title of the problem
  • Description of the problem
  • Links to documentation (E.g. papers, websites, etc...)
  • Link to source code of the executable

Definition of Done

• Code Builds without error warning
• Code is Unit Tested
• Code is Deployed to system test environment and passed system tests
• Build pushed to a demo server
• Documentation

As a researcher, I want different systems to communicate between each others with a common solution representation so that all of them can work with it and share it.

How to represent a solution to be able to work with it in different languages serializing and deserializing it efficiently.

Protocol Buffers seems to be a good option. Another option would be JSON.

As an Administrator, I want to get metrics from the execution of problems so that I can assess how the system is performing.

User Story
I need to extract metrics from both browser-based and native collaborators so I can provide insights about the execution of the problem to both researchers and collaborators as well as assess the performance of the system.

Acceptance criteria

• Get metrics from browser-based nodes
• Get metrics from native nodes
• Provide metrics to users

Definition of Done

• Code Builds without error warning
• Code is Unit Tested
• Code is Deployed to system test environment and passed system tests
• Build pushed to a demo server
• Documentation

As a deployer, I want to deploy the system in a new machine so that I can run it.

Once server is compiled, running it yields this error:

FATA[0000] Cannot open ../../../datasets/glass.csv. Error: open ../../../datasets/glass.csv: no such file or directory 

Cloud based system

Define how the system will work on an environment where you can control nodes. Like in a cloud environment where you can deal with network partitions.

Use P2P to communicate the nodes so there is no single point of failure.

As a non-tech-savvy user, I want to be able to select which project I want to contribute to so that I can collaborate in something I am interested in.

User Story
As a user of the platform that wants to contribute, I want to be able to pick a project I care about. This will encourage me to invest more time collaborating in the platform as well as sharing it with other people.

Acceptance criteria

• Show a catalogue of available problems
• By choosing one the user will be redirected to the website where they will start collaborating
• Open the problem website in a new tab
• Show Problem Title
• Show problem description
• Optionally show related documentation (e.g. Papers, Websites, etc...)
• Link to source code that will be executed on the user's browser

Definition of Done

• Code Builds without error warning
• Code is Unit Tested
• Code is Deployed to system test environment and passed system tests
• Build pushed to a demo server
• Documentation

As a researcher, I want an interface for browser-based nodes comunications so that using the distributed system is easier.

User Story
Both browser-based nodes and nodes running code natively needs to be able to share information between them in order to be synchronized. Browser-based nodes won't be able to communicate with each other due to non-existent port-forwarding on their networks that do not allow them to open ports to receive incoming connections.

Acceptance criteria

• Exchange of solutions with other peers
• Population size control
• Peers discovery
• Include inside EA library

Definition of Done

• Code Builds without error warning
• Code is Unit Tested
• Code is Deployed to system test environment and passed system tests
• Build pushed to a demo server
• Documentation

You should also test, not only train

Only TrainMLP is used in the code. You should also test.

As a researcher, I want to communicate browser-based nodes and native nodes so that both can work together towards a solution..

User Story
Both browser-based nodes and nodes running code natively needs to be able to share information between them in order to be synchronized.

The main challenge here is to address the problem of browser-based nodes not being able to communicate with each other due to non-existent port-forwarding on their networks that do not allow them to open ports to receive incoming connections.

Acceptance criteria

• Infrastructure to share solutions originated in browser-based nodes.
• Share state of available nodes for incoming requests to browser-based nodes

Definition of Done

• Code Builds without error warning
• Code is Unit Tested
• Code is Deployed to system test environment and passed system tests
• Build pushed to a demo server
• Documentation

Browser based UI

Discuss whereas to develop a system where you contribute to solving the problem by accessing an URL or using a browser extension.

Which technologies will we use to address this problem?

As a non-tech-savvy user, I want to start contributing to a problem from my browser so that I do not need to install anything on my computer.

User Story
As a user who does not know much about computers, I want to be able to contribute in a way that I do not need to install or modify anything on my machine. A web-browser enables me to run programs without installation by accessing a website.

Acceptance criteria

• A user starts collaborating to a project by accessing an URL in their browser
• The user does not need to modify anything on his machine, web-browser or network to collaborate
• Feedback is provided to the user in the browser tab that is collaborating
• Show social networks links to share the problem

Definition of Done

• Code Builds without error warning
• Code is Unit Tested
• Code is Deployed to system test environment and passed system tests
• Build pushed to a demo server
• Documentation

Cada cuanto el cliente y el servidor se comunicarán

Siempre que el cliente mejore la última mejor solución obtenida del servidor, le comunicará a este la nueva solución. El problema es, cada cuanto y como el servidor actualiza al cliente sobre la mejor solución obtenida concurrentemente?

Problems to test the platform

Which problem(s) will we use to test the performance of our platform?

Continuous Integration

Automate tests and deployment of the project.

As a researcher, I want an interface for native nodes comunications so that using the distributed system is easier.

User Story
As a researcher that wants to use the platform, I want to use it as easy as possible. I do not want to implement the underlying communication between nodes. Having an interface as if I wanted to use an already existing library for evolutionary algorithms would help me to achieve my goals faster and easier.

Acceptance criteria
The library should provide:

• Exchange of solutions between native peers
• Population size control
• Peers discovery
• Include inside EA library

Definition of Done

• Code Builds without error warning
• Code is Unit Tested
• Code is Deployed to system test environment and passed system tests
• Build pushed to a demo server
• Documentation