/msds301-wk6

Primary LanguageGo

msds301-wk6

Setup (Windows 11 + Git Bash)

  • NOTE: Executable files are included if you don't want to deal with the system dependencies
  • Clone repo with git clone git@github.com:jeremycruzz/msds301-wk6.git
  • Install GoLearn System Dependencies instructions here
    • Make sure you have g++ and MYSS2instructions here
    • Install OpenBLAS in any directory
      • git clone https://github.com/xianyi/OpenBLAS
      • cd OpenBLAS
      • make
    • Link to library
      • CGO_LDFLAGS="-L/path/to/OpenBLAS -lopenblas" go install github.com/gonum/blas/cgo

Building executables

  • Run go build -o sync.exe ./cmd/boston_linear_regression
  • Run go build -o async.exe ./cmd/boston_linear_regression_async

Running Go executable

  • Run ./sync.exe
  • Run ./async.exe

Results

Synchronous Asynchronous
Model A - Min MSE 22.466965607975983 22.22454648356535
Model A - Max MSE 32.82293524307301 35.41518392090458
Model A - Avg MSE 26.174952309117593 26.256246144534398
Model B - Min MSE 22.11739725405122 21.401751274376803
Model B - Max MSE 38.49200879493762 35.060062387146615
Model B - Avg MSE 26.589788675963973 26.518287649213583
Time - Min 2999500 ns 2076900 ns
Time - Max 5716900 ns 3343100 ns
Time - Avg 4044219 ns 2809534 ns

For the experiment, I decided to use linear regression since the median value mv is a continuous value. I removed two columns for each of the models. Model A omitted rooms and age, while Model B omitted tax and crime. Min, max, and average Mean Squared Error (MSE) were taken to ensure that both programs showed similar results over 100 runs. Additionally, from the MSE, we can see that omitting rooms and age produces a higher MSE than omitting tax and crime.

From the results, we can see that concurrency allowed the program to finish much faster (on average 43.95% faster) than the one without. The average time for the async program was faster than the minimum time for the sync program. The use of concurrency can allow us to quickly experiment with different features in linear regression models.

For future experimentation, it would be interesting to see how quickly we could test every combination of features and if doing so would produce any benefit.