Conway's Game of Life

Introduction

The British mathematician John Horton Conway devised a cellular automaton named ‘The Game of Life’. The game resides on a 2-valued 2D matrix, i.e. a binary image, where the cells can either be ‘alive’ (pixel value 255 - white) or ‘dead’ (pixel value 0 - black). The game evolution is determined by its initial state and requires no further input. Every cell interacts with its eight neighbour pixels: cells that are horizontally, vertically, or diagonally adjacent. At each matrix update in time the following transitions may occur to create the next evolution of the domain:

• any live cell with fewer than two live neighbours dies
• any live cell with two or three live neighbours is unaffected
• any live cell with more than three live neighbours dies
• any dead cell with exactly three live neighbours becomes alive

Consider the image to be on a closed domain (pixels on the top row are connected to pixels at the bottom row, pixels on the right are connected to pixels on the left and vice versa). A user can only interact with the Game of Life by creating an initial configuration and observing how it evolves. Note that evolving such complex, deterministic systems is an important application of scientific computing, often making use of parallel architectures and concurrent programs running on large computing farms.

Stage 1 - Parallel Implementation

In this stage, you are required to write code to evolve Game of Life using multiple worker goroutines on a single machine. Below are some suggested steps to help you get started. You are not required to follow them. Your implementation will be marked against the success criteria outlined below.

Step 1

Implement the Game of Life logic as it was described in the task introduction. We suggest starting with a single-threaded implementation that will serve as a starting point in subsequent steps. Your Game of Life should evolve for the number of turns specified in gol.Params.Turns. Your Game of Life should evolve the correct image specified by gol.Params.ImageWidth and gol.Params.ImageHeight.

The skeleton code starts three goroutines. The diagram below shows how they should interact with each other. Note that not all channels linking IO and the Distributor have been initialised for you. You will need to make them and add them to the distributorChannels and ioChannels structs. These structs are created in gol/gol.go.

You are not able to call methods directly on the IO goroutine. To use the IO, you will need to utilise channel communication. For reading in the initial PGM image, you will need the command, filename and input channels. Look at the file gol/io.go for details. The functions io.readPgmImage and startIo are particularly important in this step.

Your Game of Life code will interact with the user or the unit tests using the events channel. All events are defined in the file gol/event.go. In this step, you will only be working with the unit test TestGol. Therefore, you only need to send the FinalTurnComplete event.

Test your serial, single-threaded code using go test -v -run=TestGol/-1$. All the tests ran should pass.

Step 2

Parallelise your Game of Life so that it uses worker threads to calculate the new state of the board. You should implement a distributor that tasks different worker threads to operate on different parts of the image in parallel. The number of worker threads you should create is specified in gol.Params.Threads.

Note: You are free to design your system as you see fit, however, we encourage you to primarily use channels

Test your code using go test -v -run=TestGol. You can use tracing to verify the correct number of workers was used this time.

Step 3

The lab sheets included the use of a timer. Now using a ticker, report the number of cells that are still alive every 2 seconds. To report the count use the AliveCellsCount event. Also send the TurnComplete event after each complete iteration.

Test your code using go test -v -run=TestAlive.

Step 4

Implement logic to output the state of the board after all turns have completed as a PGM image.

Test your code using go test -v -run=TestPgm. Finally, run go test -v and make sure all tests are passing.

Step 5

Implement logic to visualise the state of the game using SDL. You will need to use CellFlipped and TurnComplete events to achieve this. Look at sdl/loop.go for details. Don't forget to send a CellFlipped event for all initially alive cells before processing any turns.

Also, implement the following control rules. Note that the goroutine running SDL provides you with a channel containing the relevant keypresses.

• If s is pressed, generate a PGM file with the current state of the board.
• If q is pressed, generate a PGM file with the current state of the board and then terminate the program. Your program should not continue to execute all turns set in gol.Params.Turns.
• If p is pressed, pause the processing and print the current turn that is being processed. If p is pressed again resume the processing and print "Continuing". It is not necessary for q and s to work while the execution is paused.

Test the visualisation and control rules by running go run .

Stage 2 - Distributed Implementation

In this stage, you are required to create an implementation that uses a number of AWS nodes to cooperatively calculate the new state of the Game of Life board, and communicate state between machines over a network. Below is a series of suggested steps for approaching the problem, but you are not required to follow this sequence, and can jump straight to implementing the more advanced versions of the system if you feel confident about it.

Step 1

Begin by ensuring you have a working single-threaded, single-machine implementation. You should be able to test your serial code using go test -v -run=TestGol/-1$ and all tests should pass.

Separate your implementation into two components. One component, the local controller, will be responsible for IO and capturing keypresses. The second component, the GOL Engine, will be responsible for actually processing the turns of Game of Life. You must be able to run the local controller as a client on a local machine, and the GoL engine as a server on an AWS node.

Start by implementing a basic controller which can tell the logic engine to evolve Game of Life for the number of turns specified in gol.Params.Turns. You can achieve this by implementing a single, blocking RPC call to process all requested turns.

Test your implementation using go test -v -run=TestGol/-1$ on the controller.

Step 2

You should report the number of cells that are still alive every 2 seconds to the local controller (it is anticipated that you will run a ticker on the local controller and make an RPC call to the AWS node / worker / broker every 2 seconds). The controller should then send an AliveCellsCount event to the events channel.

Test your implementation using go test -v -run=TestAlive on the controller.

Step 3

The local controller should be able to output the state of the board after all turns have completed as a PGM image.

Test your implementation using go test -v -run=TestPgm/-1$ on the controller.

Step 4

Finally, the local controller should be able to manage the behaviour of the GoL engine according to the following rules:

• If s is pressed, the controller should generate a PGM file with the current state of the board.
• If q is pressed, close the controller client program without causing an error on the GoL server. A new controller should be able to take over interaction with the GoL engine. Note that you are free to define the nature of how a new controller can take over interaction. Most likely the state will be reset. If you do manage to continue with the previous world this would be considered an extension and a form of fault tolerance.
• If k is pressed, all components of the distributed system are shut down cleanly, and the system outputs a PGM image of the latest state.
• If p is pressed, pause the processing on the AWS node and have the controller print the current turn that is being processed. If p is pressed again resume the processing and have the controller print "Continuing". It is not necessary for q and s to work while the execution is paused.

Test the control rules by running go run ..

Step 6

Reducing coupling between the "Local Controller" and the "GOL workers" is desirable. To initiate communication, the "Local Controller" connects to the broker machine via RPC. This allows the "Local Controller" to start the game by calling the main "Broker" method, which returns the final game state once it is finished. Likewise, the "Broker" connects to the "GOL workers". It is then able to give them slices of the game world and ask them to return the result of iterating on it.

Note that it is fine to have the Broker and Local Controller running on the same machine to get around firewall / port forwarding issues

Largest Image

We created a 5120x5120 pgm file if you wish to test or benchmark your solution with a very large image.

Success Criteria

• Pass all tests.
• Output the correct PGM images.
• Ensure the keyboard control rules work as needed.
• At minimum, the controller and the Game of Life engine should be separate components running on different machines (as per Step 2 above) and communicating.
• To fully satisfy the criteria your implementation should use multiple AWS nodes efficiently.

There is no need to display the live progress of the game using SDL. However, you will still need to run a blank SDL window to register the keypresses.

Extensions

Below are suggested extensions. They vary in difficulty. There are many other possible extensions to Game of Life. If you'd like to implement something that isn't an option below you're welcome to do so, but please speak to a lecturer first.

Halo Exchange

Recall that to process an iteration of Game of Life, each worker needs two extra rows (or columns). These are known as the halo regions. They need to be updated with data from neighbouring workers to process each iteration correctly. The easiest solution is to have all workers resync with a central distributor node on every iteration. This introduces a heavy communication overhead (which you might be able to measure).

Implement a Halo Exchange scheme, where workers communicate the halo regions directly to each other. Analyse the performance of your new solution and compare it with your previous implementation.

Parallel Distributed System

Add parallel workers within each distributed AWS Node.

Analyse the performance of your new solution and compare it with your previous implementation. Use various provided PGM images and analyse the effect on performance in context of the image size.

SDL Live View of Distributed Implementation

Instead of showing a blank SDL window in your local controller, add support for a Live View, in a similar way to the parallel implementation. Try to keep your implementation as efficient as possible.

Analyse the performance of your new solution and compare it with your previous implementation. Quantify and explain the overhead (if any) added by the Live View.

Fault Tolerance

Add fault tolerance to your Distributed Implementation.

In your report, explain the design of your fault tolerance mechanism. Conduct experiments to verify the effectiveness of your fault tolerance approach.

Memory Sharing

Redesign your parallel implementation to use pure memory sharing. Replace all channels with traditional synchronisation mechanisms (mutexes, sempahores, condition variables). We recommend first replacing any channels used between the workers and the distributor. Then remove channels linking the distributor with the IO and with SDL. You should still keep them as seperate goroutines. Your solution must be free of deadlocks and race conditions.

Analyse the performance of your new solution and compare it with your previous implementation. Explain any differences observed.