Serverless Simulator

A Go-based simulator for analyzing and optimizing serverless function cold starts and resource management policies, based on the paper "Serverless in the Wild: Characterizing and Optimizing the Serverless Workload at a Large Cloud Provider" (USENIX ATC '20).

Overview

This simulator implements different keep-alive policies for serverless function management and compares their effectiveness in:

• Reducing cold starts
• Optimizing memory usage
• Managing resource allocation
• Predicting function invocation patterns

The project includes implementations of:

• Fixed keep-alive policy (baseline)
• Hybrid histogram-based policy
• ARIMA-based prediction for out-of-bounds cases

Getting Started

Prerequisites

• Go 1.19 or later
• Python 3.8 or later (for visualization)
• Azure Functions dataset

Installation

1. Clone the repository:

git clone https://github.com/yourusername/serverless-simulator.git
cd serverless-simulator

2. Install Go dependencies:

go mod download

3. Install Python dependencies for visualization:

pip install -r scripts/requirements.txt

Dataset Setup

1. Download the Azure Functions dataset:

mkdir data
cd data
wget https://azurepublicdatasettraces.blob.core.windows.net/azurepublicdatasetv2/azurefunctions_dataset2019/azurefunctions-dataset2019.tar.xz
tar xf azurefunctions-dataset2019.tar.xz

Usage

Running Simulations

1. Run simulation with fixed keep-alive policy:

go run cmd/simulator/main.go \
    --policy fixed \
    --keepalive 20 \
    --data ./data \

2. Run simulation with hybrid policy:

go run cmd/simulator/main.go \
    --policy hybrid \
    --range 240 \
    --prewarm 5.0 \
    --keepalivepct 99.0 \
    --cv 2.0 \
    --data ./data \

Visualizing Results

Generate visualization plots:

python scripts/visualize.py

Project Structure

serverless-simulator/
├── cmd/                  # Command-line entry points
│   └── simulator/        # Main simulator executable
├── internal/             # Internal packages
│   ├── config/           # Configuration handling
│   ├── model/            # Core data structures
│   ├── metrics/          # Metrics for exporting data
│   ├── policy/           # Policy implementations
│   ├── simulator/        # Simulator core
│   └── dataloader/       # Dataset loading
├── pkg/                  # Public packages
│   ├── histogram/        # Histogram implementation
│   └── timeseries/       # Time series prediction
├── scripts/              # Analysis and visualization
└── data/                 # Dataset directory

Configuration Options

Simulator Flags

• --policy: Policy type (fixed or hybrid)
• --keepalive: Keep-alive period in minutes
• --range: Histogram range in minutes
• --prewarm: Pre-warm percentile
• --keepalivepct: Keep-alive percentile
• --cv: Coefficient of variation threshold
• --data: Path to data directory
• --output: Output file path

Policy Parameters

• Fixed Policy:

  • Keep-alive period (default: 20 minutes)

• Hybrid Policy:

  • Histogram range (default: 240 minutes)
  • Pre-warm percentile (default: 5.0)
  • Keep-alive percentile (default: 99.0)
  • CV threshold (default: 2.0)

Results

The simulator generates:

1. Detailed metrics for:

   • Cold start ratios
   • Memory waste time

2. Visualization plots:

   • Cold start rate
   • Wasted memory time comparison
   • Policy comparisons

Acknowledgments

• Based on research from the USENIX ATC '20 paper
• Uses the Azure Functions dataset