Link Search Menu Expand Document

Running Scalability Benchmarks

Running scalability benchmarks with Theodolite involves the following steps:

  1. Deploying a benchmark to Kubernetes
  2. Creating an execution, which describes the experimental setup for running the benchmark
  3. Accessing benchmark results
  4. Analyzing benchmark results with Theodolite’s Jupyter notebooks

Deploying a Benchmark

A benchmark specification consists of two things:

  • A Benchmark resource YAML file
  • One or more ConfigMap YAML files containing all the Kubernetes resources used by the benchmark

These files are usually provided by benchmark designers. For example, we ship Theodolite with a set of benchmarks for event-driven microservices. Alternatively, you can also create your own benchmarks.

Suppose your benchmark is defined in example-benchmark.yaml and all resources required by this benchmark are bundled in example-configmap.yaml. You can deploy both to Kubernetes by running:

kubectl apply -f example-benchmark.yaml
kubectl apply -f example-configmap.yaml

To list all benchmarks currently deployed run:

kubectl get benchmarks

The output is similar to this:

NAME                AGE   STATUS
example-benchmark   81s   Ready

The status of a benchmark tells you whether executions of it are ready to run:

  • Ready means that Theodolite has access to all resources referred from the benchmark.
  • Pending implies that not all benchmark resources are available (yet). Please ensure that you have deployed them, for example, by running kubectl get configmaps.

Creating an Execution

To run a benchmark, an Execution YAML file needs to be created such as the following one.

kind: execution
  name: theodolite-example-execution # (1) give your execution a name
  benchmark: "uc1-kstreams" # (2) refer to the benchmark to be run
    loadType: "NumSensors" # (3) chose one of the benchmark's load types
    loadValues: [25000, 50000] # (4) select a set of load intensities
    resourceType: "Instances" # (5) chose one of the benchmark's resource types
    resourceValues: [1, 2] # (6) select a set of resource amounts
  slos: # (7) set your SLOs
    - sloType: "lag trend"
      prometheusUrl: "http://prometheus-operated:9090"
      offset: 0
        threshold: 2000
        externalSloUrl: "http://localhost:80/evaluate-slope"
        warmup: 60 # in seconds
    strategy: "LinearSearch" # (8) chose a search strategy
    restrictions: ["LowerBound"] # (9) add restrictions for the strategy
    duration: 300 # (10) set the experiment duration in seconds
    repetitions: 1 # (11) set the number of repetitions
    loadGenerationDelay: 30 # (12) configure a delay before load generation
  configOverrides: []

See Creating an Execution for a more detailed explanation on how to create Executions.

Suppose your Execution resource is stored in example-execution.yaml, you can deploy it by running:

kubectl apply -f example-execution.yaml

To list all deployed executions run:

kubectl get executions

The output is similar to this:

NAME                           STATUS    DURATION   AGE
theodolite-example-execution   Running   13s        14s

The STATUS field will tell you whether a benchmark execution has been started, finished or failed due to some error. The DURATION field tells you for how long that execution is running (so far). Similar to a Kubernetes Job, an Execution is not automatically deleted once it is finished. This makes it easier to keep track of all the benchmark executions and to organize benchmark results.

Theodolite provides additional information on the current status of an Execution by producing Kubernetes events. To see them:

kubectl describe execution <execution-name>

Accessing Benchmark Results

Theodolite stores the results of benchmark executions in CSV files, whose names are starting with exp<id>_.... These files can be read and analyzed by Theodolite’s analysis notebooks.

If persisting results is enabled in Theodolite’s installation, the result files are stored in a PersistentVolume. Depending on the cluster setup or Theodolite’s configuration, the content of these volumes can usually be mounted into your host system in some way or accessed via your cloud provider.

For installations without persistence, but also as an alternative for installations with persistence, we provide a second option to access results: Theodolite comes with a results access sidecar. It allows to copy all benchmark results from the Theodolite pod to your current working directory on your host machine with the following command:

kubectl cp $(kubectl get pod -l app=theodolite -o jsonpath="{.items[0]}"):results . -c results-access

Analyzing Benchmark Results

Theodolite comes with Jupyter notebooks for analyzing and visualizing benchmark execution results. The easiest way to use them is at MyBinder:

Launch Notebooks

Alternatively, you can also run these notebook locally, for example, with Docker or Visual Studio Code.

The notebooks allow to compute a scalability function using Theodolite’s demand metric and to visualize multiple such functions in plots:

Computing the demand metric with demand-metric.ipynb (optional)

After finishing a benchmark execution, Theodolite creates a exp<id>_demand.csv file. It maps the tested load intensities to the minimal required resources for that load. If the monitoring data collected during benchmark execution should be analyzed in more detail, the demand-metric.ipynb notebook can be used.

Theodolite stores monitoring data for each conducted SLO experiment in exp<id>_<load>_<resources>_<slo-slug>_<rep>.csv files, where <id> is the ID of an execution, <load> the corresponding load intensity value, <resources> the resources value, <slo-slug> the name of the SLO and <rep> the repetition counter. The demand-metric.ipynb notebook reads these files and generates a new CSV file mapping load intensities to the minimal required resources. The format of this file corresponds to the original exp<id>_demand.csv file created when running the benchmark, but allows, for example, to evaluate different warm-up periods.

Currently, the demand-metric.ipynb notebook only supports benchmarks with the lag trend SLO out-of-the-box, but can easily be adjusted to perform any other type of analysis.

Plotting benchmark results with the demand metric with demand-metric-plot.ipynb

The demand-metric-plot.ipynb takes one or multiple exp<id>_demand.csv files as input and visualize them together in a plot. Input files can either be taken directly from Theodolite, or created from the demand-metric.ipynb notebooks.

All plotting code is only intended to serve as a template. Adjust it as needed to change colors, labels, formatting, etc. as needed. Please refer to the official docs of MatPlotLib and the ggplot style, which are used to generate the plots.

Table of contents