Skip to content

Latest commit

 

History

History
97 lines (85 loc) · 4.74 KB

README.md

File metadata and controls

97 lines (85 loc) · 4.74 KB
Raw

Iteration Support

source code: iteration.go

There are two kinds of iterations are supported. Scenario (basic iteration): iteration on application parameters for comparing results on different kind of workload Configuration: iteration on benchmark operator or node parameters for tuning application aiming at best performance

IterationSpec

Both iteration are definied in iterationSpec key of cpe.cogadvisor.io/v1/Benchmark as below.

spec:
    iterationSpec:
        iterations:
        - name: [variable name]
            values:
            - [list of values]
        configurations:
        - name: [variable name]
            values:
            - [list of values] 
        nodeSelection:
          location: [location to nodeSelector]
          values:
          - [list of tuning profile name]
          selector:
            [node label selector; matchLabels or/and matchExpressions]
            # matchLabels:
            #  label-key: label-value
            # matchExpressions:
            #  - { key: label-key, operator: <In|NotIn,Exists,DoesNotExist>, values: [label-values] }
        sequential: [true|false]
        minimize: [true|false]

When Benchmark Controller calls CreateFromOperator,

  • The iteration module will find all combinations in the list of each iteration item and create a Job for each combination with the following job name format: ` 
    [benchmark name]-cpeh-[hash32 of <iterations, build, repetition>]
  • Each item in values will be used to replace in the benchmarkSpec using Go Template.
  • For auto-tuning node, the location in nodeSelection key is used to specifiy the key location of the target node selector with dotted notation.
  • The iteration item will be also labeled to the job.
  • These labels will be sent to Parser component as a constLabels attribute.
  • constLabels will be later pushed as a label to prometheus, see output for more detail.
  • sequential is to indicate whether the iterated job should run at the same time in parallel or sequentially
  • minimize is to specify that lower number of performance value is better (default, higher is better)
  • nodeSelection key is considered as special configuration with the iteration name profile

Composite Iteration

Composite iteration referes to iteration value that is composed of more than two variable values at the same time. We support by processing the delimit ';' as an array variable. See PARSEC benchmark example.

Node Profile Tuning

tuned.go This feature is dependent on Openshift Node Tuning Operator. This will automatically check valid profile name from profiles.v1.tuned.openshift.io/rendered, label the node with each iterated value of .nodeSelection.values, then check profiles.v1.tuned.openshift.io whether the correct profile name is applied before creating a benchmark job.

  • The match label in recommend section of customized tuned resource must be in the following format. see example
  recommend:
  - match:
    - label: profile
      value: [profile name]
    priority: [integer that lower than existing tuned and unique]
    profile: [profile name]
    operand:
      debug: false

Hash Status

The hash value of iterations will be listed in the benchmark status

status:
  hash:
  - build: [buildID]
    hash: [jobHash]
    iterations:
      [iterationName]: [iteartionValue]
    run: [run number]

Results Status

When the job is completed and output is parsed and pushed as describe in output, the job tracker will also update .spec.results and .spec.bestResults.

  • .spec.results lists each iteration results.
  • .spec.bestResults presents the best performed configuration for each scenarioID derived by .iterationSpec.iterations iterations. The best performed configuration is determined by maximum performance value returned from the specified parser. In case of more than one repetition, the average value of all runs will be used.

Auto-tuning Profile

Set nodeSelection value to auto-tuned will activate node auto-tuning mechanism ; see auto-tuned Coremark benchmark

To edit node tuning search space, edit configmap cpe-operator-node-tuning-search-space and restart controller pod

kubectl edit configmap cpe-operator-node-tuning-search-space -n cpe-operator-system
kubectl delete pod $(kubectl get po -n cpe-operator-system|grep controller|tail -1|awk '{print $1}') -n cpe-operator-system