Slow Image Pull Despite Hitting Local Cache with dfdaemon

[liuyuxuan0723]

I've deployed Dragonfly in a Kubernetes cluster using the Helm chart. I've configured it to proxy a private image registry. Here are my steps:

1. On a random node, pulling a 1GB image for the first time takes 7 seconds.
2. In the var/lib/dragonfly directory of dfdaemon, I can see cached task. The monitoring shows the metrics dragonfly_scheduler_traffic with both backtosource and remotepeer traffic types.
3. After manually executing crictl rmi to remove the test image from the node and pulling again, it takes 4.5 seconds. The dfdaemon logs seem to indicate a cache hit, but the speed remains slow.

the peer log, seems to hit the local cache:

peer.txt
scheduler.txt
seed-peer.txt

My Questions:

1. Why is the pull speed slow despite hitting the cache? It doesn’t differ much from pulling directly.

My Configuration:

• containerd:

version = 2
disabled_plugins = []
imports = []
oom_score = -999
required_plugins = []
root = '/cce/containerd'
state = '/run/containerd'
[debug]
  address = '/run/containerd/debug.sock'
  level = 'info'
[plugins]
  [plugins.'io.containerd.grpc.v1.cri']
    enable_selinux = false
    enable_tls_streaming = false
    max_concurrent_downloads = 10
    sandbox_image = 'registry.baidubce.com/cce-public/pause:3.1'
    stream_server_address = '127.0.0.1'
    stream_server_port = '0'
    [plugins.'io.containerd.grpc.v1.cri'.cni]
      bin_dir = '/opt/cni/bin'
      conf_dir = '/etc/cni/net.d'
      conf_template = ''
    [plugins.'io.containerd.grpc.v1.cri'.containerd]
      default_runtime_name = 'runc'
      [plugins.'io.containerd.grpc.v1.cri'.containerd.runtimes]
        [plugins.'io.containerd.grpc.v1.cri'.containerd.runtimes.runc]
          container_annotations = []
          pod_annotations = []
          privileged_without_host_devices = false
          runtime_type = 'io.containerd.runc.v2'
    [plugins."io.containerd.grpc.v1.cri".registry]
      config_path = "/etc/containerd/certs.d"

• dfdaemon:

  aliveTime: 0s
  gcInterval: 1m0s
  keepStorage: true
  workHome:
  logDir:
  cacheDir:
  pluginDir:
  dataDir: /var/lib/dragonfly
  console: true
  health:
    path: /server/ping
    tcpListen:
      port: 40901
  verbose: false
  metrics: ":8000"
  scheduler:
    manager:
      enable: true
      netAddrs:
      - type: tcp
        addr: dragonfly-manager.dragonfly-system.svc.cluster.local:65003
      refreshInterval: 10m
    netAddrs:
    scheduleTimeout: 30s
    disableAutoBackSource: false
    seedPeer:
      clusterID: 1
      enable: false
      type: super
  host:
    idc: ""
    location: ""
  download:
    calculateDigest: false
    downloadGRPC:
      security:
        insecure: true
        tlsVerify: true
      unixListen:
        socket: ""
    peerGRPC:
      security:
        insecure: true
      tcpListen:
        port: 65000
    perPeerRateLimit: 512Mi
    prefetch: false
    totalRateLimit: 1024Mi
  upload:
    rateLimit: 1024Mi
    security:
      insecure: true
      tlsVerify: false
    tcpListen:
      port: 65002
  objectStorage:
    enable: false
    filter: Expires&Signature&ns
    maxReplicas: 3
    security:
      insecure: true
      tlsVerify: true
    tcpListen:
      port: 65004
  storage:
    diskGCThreshold: 50Gi
    multiplex: true
    strategy: io.d7y.storage.v2.simple
    taskExpireTime: 24h
  proxy:
    defaultFilter: Expires&Signature&ns
    defaultTag:
    tcpListen:
      namespace: /run/dragonfly/net
      port: 65001
    security:
      insecure: true
      tlsVerify: false
    registryMirror:
      dynamic: true
      insecure: false
      url: https://index.docker.io
    proxies:
      - regx: blobs/sha256.*
  security:
    autoIssueCert: false
    caCert: ""
    certSpec:
      validityPeriod: 4320h
    tlsPolicy: prefer
    tlsVerify: false
  network:
    enableIPv6: false
  announcer:
    schedulerInterval: 30s
  networkTopology:
    enable: false
    probe:
      interval: 20m

Environment:

• Kubernetes: 1.20
• containerd: 1.6
• Dragonfly Chart: 1.1.65

[gaius-qi]

@jim3ma