gpr-indevelopment/dissert-serverless-local-io

Source code for experiments for evaluating performance of local file system I/O workloads in serverless cloud environments.

dissert-serverless-local-io

Source code for experiments for evaluating performance of local file system I/O workloads in serverless cloud environments of AWS Lambda and Google Cloud Functions.

This repository includes the source code for functions used in the I/O experiments and for the data collection orchestrator. It also includes the R source code for data analysis alongside raw collected data.

Reproducibility package

This reproducibility package has the following pre-requisites:

1. Docker
2. RStudio

Preliminary experiments

The preliminary experiment's goal is to verify if time of day and day of week factors are statistically significant for local file system I/O workloads in AWS Lambda and Google Cloud Functions. The experiment data was saved as a CSV file. The R scripts can connect to this CSV file on local environment to produce visualization (histograms, ECDFs, etc) based on its data:

1. Open any of the R Markdown files (*.Rmd) for the preliminary experiment in RStudio.
2. Click the knit button. This will read the CSV file and generate the data visualization to a PDF file.

Main experiment

This experiment includes data from running local file system I/O workloads in AWS Lambda and Google Cloud Functions using files from 10 KB to 1 GB, with I/O sizes ranging from 512 B to 128 KB. This data was collected with minimum and maximum compatible resource allocation between these platforms.

This reproducibility package includes a custom Docker image that contains a PostgreSQL database pre-loaded with all experiment data. In addition, the R scripts can connect to this database on local environment to produce visualization (histograms, ECDFs, etc) based on its data:

1. Run the custom PostgreSQL on local using the shell command below. This will run a database named postgres on localhost:5432. Its username and password are postgres and local-db-pw, respectively.

docker run --rm -e POSTGRES_PASSWORD=local-db-pw -p 5432:5432 pimentgabriel/serverless-local-io-db

2. Then, open the R Markdown from the main experiment and knit using RStudio. This will connect to the database on local and generate the data visualization to a PDF file.

To correctly render the data visualization you might need to install the following packages to your RStudio environment:

• gridpattern
• ggpattern
• rstudioapi
• scales
• RPostgres
• ggplot2
• gridExtra
• hrbrthemes
• dplyr
• tidyr
• viridis
• readxl
• stringr

Functions

• docker-fio-lambda: Lambda function based a custom image that comes with fio benchmark installed.
• lambda-dd: Lambda function for running dd in the Lambda environment.
• gcf-dd: Google Cloud Function for running dd in the GCF environment.
• lambda-list-all-commands: Lambda function that outputs all executables on the underlying operating system of functions. Thats how we discovered dd was available.
• gcf-list-all-commands: Google Cloud Function that outputs all executables on the underlying operating system of functions. Thats how we discovered dd was available.

Other source code

• dissert-exp-orchestrator: Java Spring Boot application that periodically calls functions for collecting data. The data gets saved to a PostgreSQL database in the cloud.
• analysis: R source code and raw data used for analysis.