โšก Lazy Pulling & P2P Image Distribution

June 12, 2026

๐Ÿ’ก Lazy Pulling แ€”แ€ฟแ€ท P2P Distribution แ€€ nerdctl แ€›แ€ฒแ€ท advanced features แ€แ€ฝแ€ฑแ€•แ€ซ โ€” container startup time แ€€แ€ญแ€ฏ แ€™แ€ผแ€”แ€บแ€…แ€ฑแ€•แ€ผแ€ฎแ€ธ network bandwidth แ€žแ€€แ€บแ€žแ€ฌแ€…แ€ฑแ€แ€šแ€บแ‹

Traditional Image Pulling แ€›แ€ฒแ€ท แ€•แ€ผแ€ฟแ€”แ€ฌ

Traditional approach แ€™แ€พแ€ฌ container image pull แ€œแ€ฏแ€•แ€บแ€แ€ฒแ€ทแ€กแ€แ€ซ:

  1. Image layers แ€กแ€€แ€ฏแ€”แ€บแ€œแ€ฏแ€ถแ€ธแ€€แ€ญแ€ฏ download แ€œแ€ฏแ€•แ€บแ€›แ€แ€šแ€บ (GB แ€•แ€ฑแ€ซแ€„แ€บแ€ธแ€™แ€ปแ€ฌแ€ธแ€…แ€ฝแ€ฌ)
  2. Download แ€•แ€ผแ€ฎแ€ธแ€™แ€พ container start แ€œแ€ฏแ€•แ€บแ€œแ€ญแ€ฏแ€ทแ€›แ€แ€šแ€บ
  3. Large images แ€แ€ฝแ€ฑแ€™แ€พแ€ฌ 5-10 แ€™แ€ญแ€”แ€…แ€บ (แ€žแ€ญแ€ฏแ€ท) แ€•แ€ญแ€ฏแ€€แ€ผแ€ฌแ€แ€šแ€บ
  4. Container แ€€ image แ€›แ€ฒแ€ท 10-20% แ€œแ€ฑแ€ฌแ€€แ€บแ€žแ€ฌ แ€แ€€แ€šแ€บ แ€žแ€ฏแ€ถแ€ธแ€•แ€ฑแ€™แ€šแ€ทแ€บ 100% แ€†แ€ฝแ€ฒแ€›แ€แ€šแ€บ

Traditional Flow

docker pull myapp:latest

[========================================] 100%
Layer 1: 200MB โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ Downloaded
Layer 2: 300MB โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ Downloaded  
Layer 3: 500MB โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ Downloaded
Total: 1GB downloaded

Time: 5-10 minutes โฐ

Then container starts... ๐Ÿš€

Lazy Pulling แ€†แ€ญแ€ฏแ€แ€ฌแ€˜แ€ฌแ€œแ€ฒ?

Lazy Pulling (stargz-based) แ€€ image layers แ€แ€ฝแ€ฑแ€€แ€ญแ€ฏ on-demand แ€†แ€ฝแ€ฒแ€แ€šแ€บแ‹ Container แ€€ แ€œแ€ญแ€ฏแ€กแ€•แ€บแ€แ€ฒแ€ท files แ€แ€ฝแ€ฑแ€€แ€ญแ€ฏแ€žแ€ฌ pull แ€œแ€ฏแ€•แ€บแ€•แ€ผแ€ฎแ€ธ แ€€แ€ปแ€”แ€บแ€แ€ฌแ€แ€ฝแ€ฑแ€€แ€ญแ€ฏ แ€”แ€ฑแ€ฌแ€€แ€บแ€™แ€พ (แ€žแ€ญแ€ฏแ€ท) แ€œแ€ฏแ€ถแ€ธแ€แ€™แ€†แ€ฝแ€ฒแ€˜แ€ฐแ€ธแ‹ แ€›แ€ญแ€ฏแ€ธแ€›แ€พแ€„แ€บแ€ธแ€•แ€ผแ€ฑแ€ฌแ€›แ€„แ€บ โ€” แ€™แ€œแ€ญแ€ฏแ€กแ€•แ€บแ€žแ€ฑแ€ธแ€แ€ฌแ€แ€ฝแ€ฑ แ€™แ€†แ€ฝแ€ฒแ€˜แ€ฐแ€ธแŠ แ€œแ€ญแ€ฏแ€›แ€„แ€บ แ€†แ€ฝแ€ฒแ€แ€šแ€บ!

Lazy Pulling Flow

nerdctl run --snapshotter=stargz myapp:latest

[===>     ] 5% (50MB only!)
Downloading entrypoint + core libraries...

Container starts immediately! ๐Ÿš€โšก
Time: 5-10 seconds

(Background: Other files download as needed)

แ€˜แ€šแ€บแ€œแ€ญแ€ฏ แ€กแ€œแ€ฏแ€•แ€บแ€œแ€ฏแ€•แ€บแ€œแ€ฒ?

  1. stargz format โ€” Image layers แ€€แ€ญแ€ฏ seekable gzip (stargz) format แ€”แ€ฟแ€ท package แ€œแ€ฏแ€•แ€บแ€‘แ€ฌแ€ธแ€แ€šแ€บ
  2. TOC (Table of Contents) โ€” Layer แ€‘แ€ฒแ€™แ€พแ€ฌ file แ€แ€…แ€บแ€แ€ฏแ€แ€ปแ€„แ€บแ€ธแ€…แ€ฎ แ€˜แ€šแ€บแ€™แ€พแ€ฌแ€›แ€พแ€ญแ€œแ€ฒ index แ€›แ€พแ€ญแ€แ€šแ€บ
  3. On-demand fetch โ€” Container แ€€ file แ€แ€…แ€บแ€แ€ฏแ€€แ€ญแ€ฏ แ€–แ€แ€บแ€แ€ปแ€„แ€บแ€›แ€„แ€บ HTTP range request แ€”แ€ฟแ€ท แ€†แ€ฝแ€ฒแ€แ€šแ€บ
  4. FUSE filesystem โ€” Remote files แ€แ€ฝแ€ฑแ€€แ€ญแ€ฏ local filesystem แ€œแ€ญแ€ฏ access แ€œแ€ฏแ€•แ€บแ€œแ€ญแ€ฏแ€ทแ€›แ€กแ€ฑแ€ฌแ€„แ€บ

Lazy Pulling แ€กแ€ฌแ€ธแ€žแ€ฌแ€แ€ปแ€€แ€บแ€™แ€ปแ€ฌแ€ธ

แ€”แ€พแ€ญแ€ฏแ€„แ€บแ€ธแ€šแ€พแ€‰แ€บแ€แ€ปแ€€แ€บ (1GB Image)

Traditional Pull:
โ”œโ”€ Download: 1GB (100%)
โ”œโ”€ Time: 5-10 minutes
โ””โ”€ Container starts after download complete

Lazy Pull:
โ”œโ”€ Initial download: 50-100MB (5-10%)
โ”œโ”€ Time: 5-15 seconds
โ””โ”€ Container starts immediately
    โ””โ”€ Other files fetch on-demand (background)

Use Cases


Lazy Pulling Setup & Usage

Prerequisites

Install stargz-snapshotter

# Download latest release
wget https://github.com/containerd/stargz-snapshotter/releases/download/v0.15.1/stargz-snapshotter-v0.15.1-linux-amd64.tar.gz

# Extract
tar -C /usr/local/bin -xzf stargz-snapshotter-*.tar.gz

# Start stargz-snapshotter service
systemctl enable --now stargz-snapshotter

# Configure containerd
# Edit /etc/containerd/config.toml
[proxy_plugins]
  [proxy_plugins.stargz]
    type = "snapshot"
    address = "/run/containerd-stargz-grpc/containerd-stargz-grpc.sock"

Build stargz Image

# Method 1: nerdctl build with stargz
nerdctl build -t myapp:stargz --build-arg BUILDKIT_STEP_LOG_MAX_SIZE=-1 .
nerdctl image convert --estargz myapp:stargz myapp:lazy

# Method 2: ctr-remote convert existing image
ctr-remote image convert --estargz docker.io/library/nginx:latest nginx:lazy

# Push to registry
nerdctl push nginx:lazy

Run with Lazy Pulling

# Run container with stargz snapshotter
nerdctl run -d \
  --snapshotter=stargz \
  -p 8080:80 \
  nginx:lazy

# Check startup time
time nerdctl run --rm --snapshotter=stargz nginx:lazy echo "Started!"

P2P Image Distribution แ€†แ€ญแ€ฏแ€แ€ฌแ€˜แ€ฌแ€œแ€ฒ?

P2P (Peer-to-Peer) Distribution แ€€ container images แ€แ€ฝแ€ฑแ€€แ€ญแ€ฏ centralized registry แ€แ€…แ€บแ€แ€ฏแ€แ€Šแ€บแ€ธแ€€แ€”แ€ฑ แ€™แ€†แ€ฝแ€ฒแ€แ€ฑแ€ฌแ€ทแ€˜แ€ฒ network แ€‘แ€ฒแ€€ nodes แ€แ€ฝแ€ฑแ€€แ€ผแ€ฌแ€ธแ€™แ€พแ€ฌ แ€™แ€ปแ€พแ€แ€ฑแ€šแ€ฐแ€แ€šแ€บแ‹ BitTorrent แ€œแ€ญแ€ฏ แ€กแ€œแ€ฏแ€•แ€บแ€œแ€ฏแ€•แ€บแ€•แ€ฏแ€ถแ€•แ€ซแ€•แ€ฒแ‹ nerdctl แ€€ IPFS (InterPlanetary File System) แ€€แ€ญแ€ฏ แ€žแ€ฏแ€ถแ€ธแ€•แ€ผแ€ฎแ€ธ P2P distribution support แ€œแ€ฏแ€•แ€บแ€แ€šแ€บแ‹

Traditional vs P2P

Traditional (Centralized):

        Docker Hub / Registry
              โฌ‡๏ธ
    โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ผโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”
    โฌ‡๏ธ        โฌ‡๏ธ        โฌ‡๏ธ
  Node1    Node2    Node3
  
  Problem:
  โ€ข Registry bandwidth: 1GB ร— 100 nodes = 100GB
  โ€ข Single point of failure
  โ€ข Network bottleneck
  
P2P (Decentralized):

        Docker Hub
              โฌ‡๏ธ 1GB
           Node1 (seed)
          โฌ‡๏ธ     โฌ‡๏ธ
       Node2   Node3
         โฌ‡๏ธ  โœ–๏ธ  โฌ‡๏ธ
       Node4   Node5
       
  Benefits:
  โ€ข Registry bandwidth: 1GB only!
  โ€ข Nodes share with each other
  โ€ข Faster for many nodes
  โ€ข No single point of failure

P2P Distribution แ€กแ€ฌแ€ธแ€žแ€ฌแ€แ€ปแ€€แ€บแ€™แ€ปแ€ฌแ€ธ

แ€˜แ€šแ€บแ€œแ€ญแ€ฏ แ€กแ€œแ€ฏแ€•แ€บแ€œแ€ฏแ€•แ€บแ€œแ€ฒ? (IPFS)

  1. Content Addressing โ€” Image layers แ€€แ€ญแ€ฏ hash (CID) แ€”แ€ฟแ€ท identify แ€œแ€ฏแ€•แ€บแ€แ€šแ€บ
  2. DHT (Distributed Hash Table) โ€” แ€˜แ€šแ€บ node แ€™แ€พแ€ฌ content แ€›แ€พแ€ญแ€œแ€ฒ แ€›แ€พแ€ฌแ€แ€šแ€บ
  3. Peer Discovery โ€” Network แ€‘แ€ฒแ€€ peers แ€แ€ฝแ€ฑแ€€แ€ญแ€ฏ automatically แ€›แ€พแ€ฌแ€แ€šแ€บ
  4. Multi-source Download โ€” Peers แ€™แ€ปแ€ฌแ€ธแ€…แ€ฝแ€ฌแ€€แ€”แ€ฑ parallel แ€†แ€ฝแ€ฒแ€แ€šแ€บ
  5. Seeding โ€” Download แ€œแ€ฏแ€•แ€บแ€•แ€ผแ€ฎแ€ธ node แ€€ content แ€€แ€ญแ€ฏ แ€™แ€ปแ€พแ€แ€ฑแ€•แ€ฑแ€ธแ€แ€šแ€บ (BitTorrent แ€œแ€ญแ€ฏ)

P2P Use Cases

Real-World Scenario

Kubernetes cluster: 100 nodes, Image size: 1GB

Traditional:
โ”œโ”€ Registry โ†’ Node 1: 1GB
โ”œโ”€ Registry โ†’ Node 2: 1GB
โ”œโ”€ Registry โ†’ Node 3: 1GB
โ””โ”€ ... (97 more)
Total registry bandwidth: 100GB
Time: Serial, 5-10 minutes per node

P2P (IPFS):
โ”œโ”€ Registry โ†’ Node 1: 1GB (first node)
โ”œโ”€ Node 1 โ†’ Node 2, 3, 4: Share chunks
โ”œโ”€ Node 2, 3, 4 โ†’ Others: Further sharing
โ””โ”€ Exponential propagation
Total registry bandwidth: 1GB only!
Time: Parallel, 2-3 minutes for all nodes

๐Ÿ’ฐ Registry egress cost (AWS ECR example): $0.09/GB ร— 100GB = $9 vs $0.09 ร— 1GB = $0.09 โ€” 100x savings!


P2P Setup & Usage (IPFS)

Install IPFS

# Install IPFS (kubo)
wget https://dist.ipfs.tech/kubo/v0.25.0/kubo_v0.25.0_linux-amd64.tar.gz
tar -xvzf kubo_*.tar.gz
cd kubo
sudo bash install.sh

# Initialize IPFS
ipfs init

# Start IPFS daemon
ipfs daemon &

# Verify
ipfs version

Configure nerdctl for IPFS

# Install ipfs-registry plugin
wget https://github.com/containerd/nerdctl/releases/download/v1.7.0/nerdctl-full-1.7.0-linux-amd64.tar.gz

# IPFS แ€€ default port 5001 แ€™แ€พแ€ฌ listen แ€œแ€ฏแ€•แ€บแ€แ€šแ€บ
# nerdctl แ€€ automatic detect แ€œแ€ฏแ€•แ€บแ€แ€šแ€บ

Push Image to IPFS

# Push image to IPFS
nerdctl push ipfs://myapp:latest

# Output:
# Pushed to IPFS: QmXxx... (CID)

# Share this CID with other nodes!

Pull Image from IPFS (P2P)

# Pull from IPFS (automatically finds peers)
nerdctl pull ipfs://QmXxx...

# Or use CID directly
nerdctl run ipfs://QmXxx...

# IPFS automatically:
# 1. Finds peers with this content
# 2. Downloads from multiple peers
# 3. Becomes a seed for other peers

Lazy Pulling + P2P Together ๐Ÿš€

โšก Lazy Pulling แ€”แ€ฟแ€ท P2P แ€€แ€ญแ€ฏ แ€แ€ฝแ€ฒแ€žแ€ฏแ€ถแ€ธแ€›แ€„แ€บ แ€กแ€€แ€ฑแ€ฌแ€„แ€บแ€ธแ€†แ€ฏแ€ถแ€ธ performance แ€›แ€แ€šแ€บ โ€” แ€™แ€ผแ€”แ€บแ€แ€šแ€บแŠ bandwidth แ€žแ€€แ€บแ€žแ€ฌแ€แ€šแ€บแŠ scalable แ€–แ€ผแ€…แ€บแ€แ€šแ€บ! Combined benefits:

Example: 1000 Node Cluster

Scenario: Deploy 1GB image to 1000 nodes

Traditional:
โ”œโ”€ Download: 1GB ร— 1000 = 1TB
โ”œโ”€ Time: 10 min/node ร— 1000 = ~166 hours (serial)
โ”œโ”€ Cost: $0.09/GB ร— 1000GB = $90

Lazy + P2P:
โ”œโ”€ Download: ~100MB ร— 1 (first node) = 100MB registry
โ”‚   Rest: P2P between nodes
โ”œโ”€ Time: 5-10 seconds per node, parallel
โ”‚   Total: ~5-10 minutes for all 1000 nodes!
โ”œโ”€ Cost: $0.09/GB ร— 0.1GB = $0.009

Result: 100x faster, 10000x cheaper! ๐Ÿš€

Limitations & Considerations

Lazy Pulling Limitations

P2P Distribution Limitations


แ€™แ€พแ€แ€บแ€‘แ€ฌแ€ธแ€›แ€™แ€šแ€ทแ€บ Key Points โญ

๐ŸŽฏ แ€žแ€„แ€ทแ€บ Thesis แ€กแ€แ€ฝแ€€แ€บ: Multi-cloud orchestration แ€™แ€พแ€ฌ image distribution optimization แ€€ แ€กแ€›แ€ฑแ€ธแ€€แ€ผแ€ฎแ€ธแ€แ€šแ€บแ‹ Lazy pulling แ€”แ€ฟแ€ท P2P แ€€ cross-cloud deployments แ€™แ€พแ€ฌ startup time แ€”แ€ฟแ€ท bandwidth cost แ€€แ€ญแ€ฏ แ€žแ€ญแ€žแ€ญแ€žแ€ฌแ€žแ€ฌ แ€œแ€ปแ€พแ€ฑแ€ฌแ€ทแ€แ€ปแ€•แ€ฑแ€ธแ€”แ€ญแ€ฏแ€„แ€บแ€แ€šแ€บแ‹ NSGA-II แ€€ placement decide แ€œแ€ฏแ€•แ€บแ€•แ€ผแ€ฎแ€ธแ€›แ€„แ€บ nerdctl แ€€ efficient distribution แ€œแ€ฏแ€•แ€บแ€•แ€ฑแ€ธแ€™แ€šแ€บ!