Skip to content

Benchmarks_g5k.md

Latest commit

 

History

History
62 lines (47 loc) · 3.62 KB

Benchmarks_g5k.md

File metadata and controls

62 lines (47 loc) · 3.62 KB

Running benchmarks

Motivation benchmarks

Open the ./notebooks/motivation/xp.ipynb notebook in your browser (make sure your jupyter server is still up and running, following the Requirements_g5k.md instruction).

Before continuing, make sure that the Flink image name in the following file is the same as the one you used during the build phase:

apiVersion: flink.apache.org/v1beta1
kind: FlinkDeployment
metadata:
  name: flink
spec:
  image: flink-justin:dais <------- Your Flink image name
  1. ./notebooks/motivation/read-only/query.yaml
  2. ./notebooks/motivation/write-only/query.yaml
  3. ./notebooks/motivation/update/query.yaml

Follow the instructions in the notebook to execute the motivation benchmarks.

⚠️ Grid5000 only allows access to the site from the front end or with the use of a VPN. This VPN can be easily configured through the website. With the help of ingresses, the services are available through the following ports:

  1. Flink: node.id.site:8081
  2. Grafana: node.id.site:30300

Nexmark benchmarks

Open the ./notebooks/nexmark/xp.ipynb notebook in your browser (make sure your jupyter server is still up and running, following the Requirements.md instruction).

Before continuing, make sure that the Flink image name in the following file is the same as the one you used during the build phase:

apiVersion: flink.apache.org/v1beta1
kind: FlinkDeployment
metadata:
  name: flink
spec:
  image: flink-justin:dais <------- Your Flink image name
  1. ./notebooks/motivation/read-only/query.yaml, with X being the query number.

The notebook contains a cell for each query. These cells will execute the query twice, one with the default auto scaler, and one with the Justin auto scaler.

Playing with Justin

As explained in the paper, Justin relies on multiple parameters such as a minimum cache hit rate and minimum state access latency. These parameters can be modified via the YAML file of the query, namely the lines:

  1. job.autoscaler.cache-hit-rate.min.threshold: "0.8" (ratio)
  2. job.autoscaler.state-latency.threshold: "1000000.0" (nano seconds)

Note that changing those values can result in different scaling decisions.

Additionally, the reader can also modify the stabilization interval (job.autoscaler.stabilization.interval) and metric window (job.autoscaler.metrics.window).

Modifying the Justin Policy.

The Justin policy is implemented in the Flink Kubernetes Operator's auto scaler module. The reader can easily apply its own policy logic by modifying the policy method of the ScalingExecutor.java file. The reviewer has access to previous Scaling Decisions (if any) through the scaling parameter.

Once modified, the image needs to be rebuilt and pushed to the nodes. Delete the operator and its resources by executing the script delete.sh located in the scriptsfolder. This script will delete any Flink job pending, the operator, and the 3 custom resource definitions. Once deleted, you can re-deploy the operator following the previous instructions.