Aria2 & Multi-Threaded Concurrency: Conquering GitHub Release Download Bottlenecks

1. Why Single-Threaded HTTP Downloads Stall on Large Releases

When downloading multi-gigabyte Release assets (such as PyTorch wheels, LLM weights, or CUDA toolchains) through a standard web browser or single-stream curl, transfer rates frequently languish at 50~200 KB/s. This is primarily governed by the Bandwidth-Delay Product (BDP):

$$\text{BDP} = \text{Bandwidth} \times \text{Round-Trip Time (RTT)}$$

Across trans-Pacific routes where RTT reaches 250ms, a single TCP socket’s TCP Window Scale buffer must remain flawlessly saturated without packet drops to maintain speed. The moment a 2% packet loss occurs, TCP’s congestion avoidance algorithm slashes throughput.

By contrast, Multi-Threaded Concurrent Slicing partitions a file into multiple byte ranges, opening 16 parallel TCP connections simultaneously. Even if one stream experiences a brief packet drop, the other 15 connections continue transferring at line rate.

2. Multi-Threaded Slicing & HTTP 206 Range

GitHub Acceleration Doctor passes through HTTP Range: bytes=start-end headers transparently to upstream AWS S3 buckets. When an Aria2 client requests chunks:

[Aria2 Downloader]
  ├── Socket 1: Range bytes=0-10485759      ──> [ghproxy.icu] ──> [S3 Partial Content 206]
  ├── Socket 2: Range bytes=10485760-20971519 ─> [ghproxy.icu] ──> [S3 Partial Content 206]
  ├── ...
  └── Socket 16: Range bytes=157286400-...    ─> [ghproxy.icu] ──> [S3 Partial Content 206]

Each socket downloads its independent byte slice into a pre-allocated sparse file, stitching chunks together seamlessly upon completion.

3. High-Performance Aria2 CLI Commands

Install Aria2 on your operating system:

# Ubuntu / Debian
sudo apt-get install -y aria2

# macOS
brew install aria2

# CentOS / RHEL
sudo yum install -y aria2

Execute accelerated 16-thread downloads:

aria2c -s 16 -x 16 -k 1M -c "https://ghproxy.icu/https://github.com/ollama/ollama/releases/download/v0.5.1/ollama-linux-amd64.tgz"

Key Parameter Explanations:

• -s 16: Split the file into 16 parts;
• -x 16: Allow up to 16 concurrent TCP connections per host;
• -k 1M: Minimum split slice size set to 1MB;
• -c: Enable auto-resumption (continue), allowing interrupted downloads to pick up exactly where they stopped without discarding downloaded data.

4. Automation Wrapper Function for .bashrc or .zshrc

To make downloading effortless in your daily workflow, add the following helper function to ~/.bashrc or ~/.zshrc:

# Rapid GitHub Release Downloader
ghget() {
  if [ -z "$1" ]; then
    echo "Usage: ghget <github-url>"
    return 1
  fi
  local target_url="$1"
  if [[ "$target_url" =~ ^https?:// ]]; then
    local proxy_url="https://ghproxy.icu/${target_url}"
  else
    local proxy_url="https://ghproxy.icu/https://${target_url}"
  fi
  echo ">>> Downloading via GitHub Acceleration Doctor..."
  aria2c -s 16 -x 16 -k 1M -c "$proxy_url"
}

Now you can simply run:

ghget https://github.com/cli/cli/releases/download/v2.45.0/gh_2.45.0_linux_amd64.tar.gz

5. Performance Comparison: Single Stream vs. 16-Thread Concurrency