Quick links

• Repository structure
• Algorithm
• Deployment
• Walkthrough
• Implementation
• Known Issues

Using Kubernetes

In case your are using Kubernetes through Kind and this is installed remotely, you need to follow these steps:

1. create a cluster with kind
2. let kubectl proxy running in your remote node
3. create a ssh-tunnel to your remote node at port 8001: ssh -N -L 8001:127.0.0.1:8001
4. access any service running in kubernetes using this url template: http://localhost:8001/api/v1/namespaces/<namespace>/services/<service-name>:8081/proxy/. For more details refer to https://kubernetes.io/docs/tasks/access-application-cluster/access-cluster/

Build

Run make build in SmartTuning directory.

Applications tested

• AcmeAir
• Daytrader
• Quarkus Demo

Release notes

Version 4.0

• smarttuning architecture relies on a state machine to progress over different tuning stages
• tuning multi-replicas services #21
• classify workloads based on the number of replicas
• classify workloads based on throughput
• add mock workload-classifier
• remove proxy need and service replication

Version 3.0

• update codebase to create one tuning context per workload #18
  • add eager stop
  • add option to turn on/off eager stop on poor configs
  • add pruning mechanism to eager stop a configuration sampled to another worklaod
  • add scheduler to switch tuning contexts and workload changes
• update codebase to use Optuna #17
• add new metrics to be used on objective function #d2de64, #5d8deb
• change trigger for restarting replicas to use mean+stddev rather median #67ff0c

Version 2.0

• ability to tune multiple pods at once
• quick-abortion of poor configs #1
• set boundaries of dependent parameters on-the-fly #10
• uses two-phase algorithm to update a pod #13