# DEMO - Single-Machine Multiple-Dataset Speed Up Accessing Remote Files
Powered by Alluxio (opens new window) and Fuse (opens new window), Fluid provides a simple way for users to access files stored in remote filesystems, just like accessing some ordinary files in local filesystem. Fluid manages and isolates the entire life cycle of data sets, especially for short life cycle applications (e.g data analysis tasks, machine learning tasks), users can deploy them on a large scale in a cluster.
This demo aims to show you an overview of all the features mentioned above.
# Prerequisites
Before everything we are going to do, please refer to Installation Guide (opens new window) to install Fluid on your Kubernetes Cluster, and make sure all the components used by Fluid are ready like this:
$ kubectl get pod -n fluid-system
NAME READY STATUS RESTARTS AGE
alluxioruntime-controller-5b64fdbbb-84pc6 1/1 Running 0 8h
csi-nodeplugin-fluid-fwgjh 2/2 Running 0 8h
csi-nodeplugin-fluid-ll8bq 2/2 Running 0 8h
dataset-controller-5b7848dbbb-n44dj 1/1 Running 0 8h
Normally, you shall see a Pod named controller-manager and several Pods named csi-nodeplugin. The num of csi-nodeplugin Pods depends on how many nodes your Kubernetes cluster have, so please make sure all csi-nodeplugin Pods are working properly.
# Install Resources to Kubernetes
Label a node
$ kubectl label node cn-beijing.192.168.0.199 fluid=multi-dataset
In the next steps, we will use
NodeSelectorto manage the nodes scheduled by the Dataset. There, it is only for experiment use.
Check the Dataset object to be created
$ cat<<EOF >dataset.yaml
apiVersion: data.fluid.io/v1alpha1
kind: Dataset
metadata:
name: hbase
spec:
mounts:
- mountPoint: https://mirrors.tuna.tsinghua.edu.cn/apache/hbase/stable/
name: hbase
nodeAffinity:
required:
nodeSelectorTerms:
- matchExpressions:
- key: fluid
operator: In
values:
- "multi-dataset"
placement: "Shared" // set Exclusive or empty means dataset exclusive
EOF
$ cat<<EOF >dataset1.yaml
apiVersion: data.fluid.io/v1alpha1
kind: Dataset
metadata:
name: spark
spec:
mounts:
- mountPoint: https://mirrors.bit.edu.cn/apache/spark/
name: spark
nodeAffinity:
required:
nodeSelectorTerms:
- matchExpressions:
- key: fluid
operator: In
values:
- "multi-dataset"
placement: "Shared"
EOF
Notes: Here, we use THU's tuna Apache mirror site as our
mountPoint. If your environment isn't in Chinese mainland, please replace it withhttps://downloads.apache.org/hbase/stable/andhttps://downloads.apache.org/spark/.
Here, we'd like to create a resource object with kind Dataset. Dataset is a Custom Resource Definition(CRD) defined by Fluid and used to tell Fluid where to find all the data you'd like to access. Under the hood, Fluid uses Alluxio to do some mount operations, so mountPoint property can be any legal UFS path acknowledged by Alluxio. Here, we use WebUFS (opens new window) for its simplicity.
For more information about UFS, please refer to Alluxio Docs - Storage Integrations (opens new window).
We use Hbase stable and Spark on a mirror site of Apache downloads as an example of remote file. It's nothing special, you can change it to any remote file you like. But please note that, if you are going to use WebUFS like we do, files on Apache sites are highly recommended because you might need some advanced configurations (opens new window) due to current implementation of WebUFS.
Create the Dataset object
$ kubectl apply -f dataset.yaml
dataset.data.fluid.io/hbase created
$ kubectl apply -f dataset1.yaml
dataset.data.fluid.io/spark created
Check status of the Dataset object
$ kubectl get dataset
NAME UFS TOTAL SIZE CACHED CACHE CAPACITY CACHED PERCENTAGE PHASE AGE
hbase NotBound 6s
spark NotBound 4s
With a NotBound phase in status, the dataset is not ready cause there isn't any AlluxioRuntime object supporting it. We'll create one in the following steps.
Check the AlluxioRuntime object to be created
$ cat<<EOF >runtime.yaml
apiVersion: data.fluid.io/v1alpha1
kind: AlluxioRuntime
metadata:
name: hbase
spec:
replicas: 1
tieredstore:
levels:
- mediumtype: MEM
path: /dev/shm
quota: 2Gi
high: "0.95"
low: "0.7"
EOF
$ cat<<EOF >runtime1.yaml
apiVersion: data.fluid.io/v1alpha1
kind: AlluxioRuntime
metadata:
name: spark
spec:
replicas: 1
tieredstore:
levels:
- mediumtype: MEM
path: /dev/shm
quota: 4Gi
high: "0.95"
low: "0.7"
EOF
Create a AlluxioRuntime object
$ kubectl create -f runtime.yaml
alluxioruntime.data.fluid.io/hbase created
# Pay attention to waiting for all components of Dataset hbase Running
$ kubectl get pod -o wide | grep hbase
NAME READY STATUS RESTARTS AGE IP NODE NOMINATED NODE READINESS GATES
hhbase-fuse-jl2g2 1/1 Running 0 2m24s 192.168.0.199 cn-beijing.192.168.0.199 <none> <none>
hbase-master-0 2/2 Running 0 2m55s 192.168.0.200 cn-beijing.192.168.0.200 <none> <none>
hbase-worker-g89p8 2/2 Running 0 2m24s 192.168.0.199 cn-beijing.192.168.0.199 <none> <none>
$ kubectl create -f runtime1.yaml
alluxioruntime.data.fluid.io/spark created
Get the AlluxioRuntime object
$ kubectl get alluxioruntime
NAME MASTER PHASE WORKER PHASE FUSE PHASE AGE
hbase Ready Ready Ready 2m14s
spark Ready Ready Ready 58s
AlluxioRuntime is another CRD defined by Fluid. An AluxioRuntime object describes specifications used to run an Alluxio instance.
Wait for a while, and make sure all components defined in the AlluxioRuntime object are ready. You shall see something like this:
$ kubectl get pod -o wide
NAME READY STATUS RESTARTS AGE IP NODE NOMINATED NODE READINESS GATES
hhbase-fuse-jl2g2 1/1 Running 0 2m24s 192.168.0.199 cn-beijing.192.168.0.199 <none> <none>
hbase-master-0 2/2 Running 0 2m55s 192.168.0.200 cn-beijing.192.168.0.200 <none> <none>
hbase-worker-g89p8 2/2 Running 0 2m24s 192.168.0.199 cn-beijing.192.168.0.199 <none> <none>
spark-fuse-5z49p 1/1 Running 0 19s 192.168.0.199 cn-beijing.192.168.0.199 <none> <none>
spark-master-0 2/2 Running 0 50s 192.168.0.200 cn-beijing.192.168.0.200 <none> <none>
spark-worker-96ksn 2/2 Running 0 19s 192.168.0.199 cn-beijing.192.168.0.199 <none> <none>
Note that the worker and fuse components of the different Datasets above can be dispatched to the same node cn-beijing.192.168.0.199 normally .
Check status of the Dataset object again
$ kubectl get dataset
NAME UFS TOTAL SIZE CACHED CACHE CAPACITY CACHED PERCENTAGE PHASE AGE
hbase 443.89MiB 0.00B 2.00GiB 0.0% Bound 11m
spark 1.92GiB 0.00B 4.00GiB 0.0% Bound 9m38s
Because it has been bound to a successfully started AlluxioRuntime, the state of the Dataset resource object has been updated, and the value of the PHASE attribute has changed to the Bound state. The basic information about the resource object can be obtained through the above command.
Check status of the AlluxioRuntime object
$ kubectl get alluxioruntime -o wide
NAME READY MASTERS DESIRED MASTERS MASTER PHASE READY WORKERS DESIRED WORKERS WORKER PHASE READY FUSES DESIRED FUSES FUSE PHASE AGE
hbase 1 1 Ready 1 1 Ready 1 1 Ready 11m
spark 1 1 Ready 1 1 Ready 1 1 Ready 9m52s
Detailed information about the Alluxio instance is provided here.
Check related PersistentVolume and PersistentVolumeClaim
$ kubectl get pv
NAME CAPACITY ACCESS MODES RECLAIM POLICY STATUS CLAIM STORAGECLASS REASON AGE
hbase 100Gi RWX Retain Bound default/hbase 4m55s
spark 100Gi RWX Retain Bound default/spark 51s
$ kubectl get pvc
NAME STATUS VOLUME CAPACITY ACCESS MODES STORAGECLASS AGE
hbase Bound hbase 100Gi RWX 4m57s
spark Bound spark 100Gi RWX 53s
Related PV and PVC have been created by Fluid since the Dataset object is ready(bounded). Workloads are now able to access remote files by mounting PVC.
# Remote File Access
Check the app to be created
$ cat<<EOF >nginx.yaml
apiVersion: v1
kind: Pod
metadata:
name: nginx-hbase
spec:
containers:
- name: nginx
image: nginx
volumeMounts:
- mountPath: /data
name: hbase-vol
volumes:
- name: hbase-vol
persistentVolumeClaim:
claimName: hbase
nodeName: cn-beijing.192.168.0.199
EOF
$ cat<<EOF >nginx1.yaml
apiVersion: v1
kind: Pod
metadata:
name: nginx-spark
spec:
containers:
- name: nginx
image: nginx
volumeMounts:
- mountPath: /data
name: hbase-vol
volumes:
- name: hbase-vol
persistentVolumeClaim:
claimName: spark
nodeName: cn-beijing.192.168.0.199
EOF
Run a demo app to access remote files
$ kubectl create -f nginx.yaml
$ kubectl create -f nginx1.yaml
Login to nginx hbase Pod:
$ kubectl exec -it nginx-hbase -- bash
Check file status:
$ ls -lh /data/hbase
total 444M
-r--r----- 1 root root 193K Sep 16 00:53 CHANGES.md
-r--r----- 1 root root 112K Sep 16 00:53 RELEASENOTES.md
-r--r----- 1 root root 26K Sep 16 00:53 api_compare_2.2.6RC2_to_2.2.5.html
-r--r----- 1 root root 211M Sep 16 00:53 hbase-2.2.6-bin.tar.gz
-r--r----- 1 root root 200M Sep 16 00:53 hbase-2.2.6-client-bin.tar.gz
-r--r----- 1 root root 34M Sep 16 00:53 hbase-2.2.6-src.tar.gz
Login to nginx spark Pod:
$ kubectl exec -it nginx-spark -- bash
Check file status:
$ ls -lh /data/spark/
total 1.0K
dr--r----- 1 root root 7 Oct 22 12:21 spark-2.4.7
dr--r----- 1 root root 7 Oct 22 12:21 spark-3.0.1
$ du -h /data/spark/
999M /data/spark/spark-3.0.1
968M /data/spark/spark-2.4.7
2.0G /data/spark/
Loginout nginx Pod:
$ exit
As you may have seen, all the files on the WebUFS appear no differences from any other file in the local filesystem of the nginx Pod.
# Speed Up Accessing Remote Files
To demonstrate how great speedup you may enjoy when accessing remote files, here is a demo job:
Check the test job to be launched
$ cat<<EOF >app.yaml
apiVersion: batch/v1
kind: Job
metadata:
name: fluid-copy-test-hbase
spec:
template:
spec:
restartPolicy: OnFailure
containers:
- name: busybox
image: busybox
command: ["/bin/sh"]
args: ["-c", "set -x; time cp -r /data/hbase ./"]
volumeMounts:
- mountPath: /data
name: hbase-vol
volumes:
- name: hbase-vol
persistentVolumeClaim:
claimName: hbase
nodeName: cn-beijing.192.168.0.199
EOF
$ cat<<EOF >app1.yaml
apiVersion: batch/v1
kind: Job
metadata:
name: fluid-copy-test-spark
spec:
template:
spec:
restartPolicy: OnFailure
containers:
- name: busybox
image: busybox
command: ["/bin/sh"]
args: ["-c", "set -x; time cp -r /data/spark ./"]
volumeMounts:
- mountPath: /data
name: spark-vol
volumes:
- name: spark-vol
persistentVolumeClaim:
claimName: spark
nodeName: cn-beijing.192.168.0.199
EOF
Launch a test job
$ kubectl create -f app.yaml
job.batch/fluid-copy-test-hbase created
$ kubectl create -f app1.yaml
job.batch/fluid-copy-test-spark created
The hbase job executes a shell command time cp -r /data/hbase ./, and /data/hbase is the location where the remote file is mounted in the pod. After the command is completed, the execution time of the command will be displayed on the terminal.
The spark job executes a shell command time cp -r /data/spark ./, and /data/hbase is the location where the remote file is mounted in the pod. After the command is completed, the execution time of the command will be displayed on the terminal.
Wait for a while and make sure the job has completed. You can check its runnning status by:
$ kubectl get pod -o wide | grep copy
ffluid-copy-test-hbase-r8gxp 0/1 Completed 0 4m16s 172.29.0.135 cn-beijing.192.168.0.199 <none> <none>
fluid-copy-test-spark-54q8m 0/1 Completed 0 4m14s 172.29.0.136 cn-beijing.192.168.0.199 <none> <none>
If you see the above result, it means the job has been completed.
Note:
r8gxpinfluid-copy-test-hbase-r8gxpis a specifier generated by the Job we created. It's highly possible that you may have different specifier in your environment. Please remember replace it with your own specifier in the following steps
Check running time of the test job
$ kubectl logs fluid-copy-test-hbase-r8gxp
+ time cp -r /data/hbase ./
real 3m 34.08s
user 0m 0.00s
sys 0m 1.24s
$ kubectl logs fluid-copy-test-spark-54q8m
+ time cp -r /data/spark ./
real 3m 25.47s
user 0m 0.00s
sys 0m 5.48s
It can be seen that the first remote file read hbase took nearly 3m34s, and the spark took nearly 3m25s.
Check status of the dataset
$ kubectl get dataset
NAME UFS TOTAL SIZE CACHED CACHE CAPACITY CACHED PERCENTAGE PHASE AGE
hbase 443.89MiB 443.89MiB 2.00GiB 100.0% Bound 30m
spark 1.92GiB 1.92GiB 4.00GiB 100.0% Bound 28m
Now, all the remote files have been cached in Alluxio.
Re-Launch the test job
$ kubectl delete -f app.yaml
$ kubectl create -f app.yaml
$ kubectl delete -f app1.yaml
$ kubectl create -f app1.yaml
Since the remote file has been cached, the test job can be completed quickly:
$ kubectl get pod -o wide| grep fluid
fluid-copy-test-hbase-sf5md 0/1 Completed 0 53s 172.29.0.137 cn-beijing.192.168.0.199 <none> <none>
fluid-copy-test-spark-fwp57 0/1 Completed 0 51s 172.29.0.138 cn-beijing.192.168.0.199 <none> <none>
$ kubectl logs fluid-copy-test-hbase-sf5md
+ time cp -r /data/hbase ./
real 0m 0.36s
user 0m 0.00s
sys 0m 0.36s
$ kubectl logs fluid-copy-test-spark-fwp57
+ time cp -r /data/spark ./
real 0m 1.57s
user 0m 0.00s
sys 0m 1.57s
Doing the same read operation, hbase takes only 0.36s this time and spark takes only 1.57s.
The great speedup attributes to the powerful caching capability provided by Alluxio. That means that once you access some remote file, it will be cached in Alluxio, and your next following operations will enjoy a local access instead of a remote one, and thus a great speedup.
Note: Time spent for the test job depends on your network environment. If it takes too long for you to complete the job, changing a mirror or some smaller file might help.
Also login to the host node (if possible)
$ ssh root@192.168.0.199
$ ls /dev/shm/default/
hbase spark
$ ls -lh /dev/shm/default/hbase/alluxioworker/
total 444M
-rwxrwxrwx 1 root root 174K Oct 22 20:27 100663296
-rwxrwxrwx 1 root root 115K Oct 22 20:27 16777216
-rwxrwxrwx 1 root root 200M Oct 22 20:26 33554432
-rwxrwxrwx 1 root root 106K Oct 22 20:26 50331648
-rwxrwxrwx 1 root root 211M Oct 22 20:27 67108864
-rwxrwxrwx 1 root root 34M Oct 22 20:27 83886080
$ ls -lh /dev/shm/default/spark/alluxioworker/
total 2.0G
-rwxrwxrwx 1 root root 210M Oct 22 21:06 100663296
-rwxrwxrwx 1 root root 16M Oct 22 21:07 117440512
-rwxrwxrwx 1 root root 195M Oct 22 21:05 134217728
-rwxrwxrwx 1 root root 214M Oct 22 21:05 150994944
-rwxrwxrwx 1 root root 140M Oct 22 21:08 16777216
-rwxrwxrwx 1 root root 22M Oct 22 21:05 167772160
-rwxrwxrwx 1 root root 221M Oct 22 21:07 184549376
-rwxrwxrwx 1 root root 150M Oct 22 21:06 201326592
-rwxrwxrwx 1 root root 311K Oct 22 21:07 218103808
-rwxrwxrwx 1 root root 322K Oct 22 21:06 234881024
-rwxrwxrwx 1 root root 210M Oct 22 21:06 33554432
-rwxrwxrwx 1 root root 161M Oct 22 21:07 50331648
-rwxrwxrwx 1 root root 223M Oct 22 21:07 67108864
-rwxrwxrwx 1 root root 208M Oct 22 21:07 83886080
It can be seen that the block files cached by different Datasets are isolated according to the namespace and name of the Dataset.
# Clean Up
$ kubectl delete -f .
$ kubectl label node cn-beijing.192.168.0.199 fluid-