EdgeFS Scale-Out NFS CRD

Rook allows creation and customization of EdgeFS NFS file systems through the custom resource definitions (CRDs). The following settings are available for customization of EdgeFS NFS services.

Sample

  1. apiVersion: edgefs.rook.io/v1beta1
  2. kind: NFS
  3. metadata:
  4.   name: nfs01
  5.   namespace: rook-edgefs
  6. spec:
  7.   instances: 3
  8.   #relaxedDirUpdates: true
  9.   #chunkCacheSize: 1Gi
  10.   # A key/value list of annotations
  11.   annotations:
  12.   #  key: value
  13.   placement:
  14.   #  nodeAffinity:
  15.   #    requiredDuringSchedulingIgnoredDuringExecution:
  16.   #      nodeSelectorTerms:
  17.   #      - matchExpressions:
  18.   #        - key: role
  19.   #          operator: In
  20.   #          values:
  21.   #          - nfs-node
  22.   #  tolerations:
  23.   #  - key: nfs-node
  24.   #    operator: Exists
  25.   #  podAffinity:
  26.   #  podAntiAffinity:
  27.   #resourceProfile: embedded
  28.   resources:
  29.   #  limits:
  30.   #    cpu: "500m"
  31.   #    memory: "1024Mi"
  32.   #  requests:
  33.   #    cpu: "500m"
  34.   #    memory: "1024Mi"

Metadata

  • name: The name of the NFS system to create, which must match existing EdgeFS service.
  • namespace: The namespace of the Rook cluster where the NFS service is created.
  • instances: The number of active NFS service instances. EdgeFS NFS service is Multi-Head capable, such so that multiple PODs can mount same tenant’s buckets via different endpoints. EdgeFS CSI provisioner orchestrates distribution and load balancing across NFS service instances in round-robin or random policy ways.
  • relaxedDirUpdates: If set to true then it will significantly improve performance of directory operations by defering updates, guaranteeing eventual directory consistency. This option is recommended when a bucket exported via single NFS instance and it is not a destination for ISGW Link synchronization.
  • chunkCacheSize: Limit amount of memory allocated for dynamic chunk cache. By default NFS pod uses up to 75% of available memory as chunk caching area. This option can influence this allocation strategy.
  • annotations: Key value pair list of annotations to add.
  • placement: The NFS PODs can be given standard Kubernetes placement restrictions with nodeAffinity, tolerations, podAffinity, and podAntiAffinity similar to placement defined for daemons configured by the cluster CRD.
  • resourceProfile: NFS pod resource utilization profile (Memory and CPU). Can be embedded or performance (default). In case of performance an NFS pod trying to increase amount of internal I/O resources that results in higher performance at the cost of additional memory allocation and more CPU load. In embedded profile case, NFS pod gives preference to preserving memory over I/O and limiting chunk cache (see chunkCacheSize option). The performance profile is the default unless cluster wide embedded option is defined.
  • resources: Set resource requests/limits for the NFS Pod(s), see Resource Requirements/Limits.

Setting up EdgeFS namespace and tenant

For more detailed instructions please refer to EdgeFS Wiki.

Below is an exampmle procedure to get things initialized and configured.

Before new local namespace (or local site) can be used, it has to be initialized with FlexHash and special purpose root object.

FlexHash consists of dynamically discovered configuration and checkpoint of accepted distribution table. FlexHash is responsible for I/O direction and plays important role in dynamic load balancing logic. It defines so-called Negotiating Groups (typically across zoned 8-24 disks) and final table distribution across all the participating components, e.g. data nodes, service gateways and tools.

Root object holds system information and table of namespaces registered to a local site. Root object is always local and never shared between the sites.

To initialize system and prepare logical definitions, login to the toolbox as shown in this example:

  1. kubectl get po --all-namespaces | grep edgefs-mgr
  2. kubectl exec -it -n rook-edgefs rook-edgefs-mgr-6cb9598469-czr7p -- env COLUMNS=$COLUMNS LINES=$LINES TERM=linux toolbox

Assumption at this point is that nodes are all configured and can be seen via the following command:

  1. efscli system status
  1. Initialize cluster

Verify that HW (or better say emulated in this case) configuration look normal and accept it

  1. efscli system init

At this point new dynamically discovered configuration checkpoint will be created at $NEDGE_HOME/var/run/flexhash-checkpoint.json This will also create system “root” object, holding Site’s Namespace. Namespace may consist of more then single region.

  1. Create new local namespace (or we also call it “Region” or “Segment”)
  1. efscli cluster create Hawaii
  1. Create logical tenants of cluster namespace “Hawaii”, also buckets if needed
  1. efscli tenant create Hawaii/Cola
  2. efscli bucket create Hawaii/Cola/bk1
  3. efscli tenant create Hawaii/Pepsi
  4. efscli bucket create Hawaii/Pepsi/bk1

Now cluster is setup, services can be now created and attached to CSI provisioner.

  1. Create NFS service objects for tenants
  1. efscli service create nfs nfs-cola
  2. efscli service serve nfs-cola Hawaii/Cola/bk1
  3. efscli service create nfs nfs-pepsi
  4. efscli service serve nfs-pepsi Hawaii/Pepsi/bk1
  1. Create NFS CRDs
  1. apiVersion: edgefs.rook.io/v1beta1
  2. kind: NFS
  3. metadata:
  4.   name: nfs-cola
  5.   namespace: rook-edgefs
  6. spec:
  7.   instances: 1
  1. apiVersion: edgefs.rook.io/v1beta1
  2. kind: NFS
  3. metadata:
  4.   name: nfs-pepsi
  5.   namespace: rook-edgefs
  6. spec:
  7.   instances: 1

At this point two NFS services should be available. Verify that showmount command can see service (substitute CLUSTERIP with corresponding entry from kubectl get svc command):

  1. kubectl get svc --all-namespaces
  2. showmount -e CLUSTERIP