Workflow

This page describes the available parameters for defining workflows in Daisytuner. Workflows control how and when your benchmarks are executed on our cluster, daisyfield.

The on Section

This section specifies conditions that trigger the workflow. The syntax closely follows GitHub Actions workflows:

Parameter	Values	Description
push	branches with a list of branch names, or tags with a list of tags	Specifies workflow triggers upon pushes to specific branches or tags. Syntax is identical to GitHub Actions’ push events.
pull_request	types (e.g., opened, synchronize, ready_for_review)	Specifies workflow triggers for pull request events. Available event types match GitHub Actions pull request triggers.

Example:

on:
  push:
    branches:
      - main
      - develop
  pull_request:
    types: [opened, reopened, synchronize]

The parameters Section

The parameters section defines job-specific execution settings:

Parameter	Type	Description
partitions	List of strings	Specifies which cluster partitions your benchmarks should run on. See the Partitions Reference for available options.
timeout (optional)	Integer	Maximum allowed execution time in minutes before the benchmark job is automatically terminated. The default value is 20 minutes.

parameters:
  partitions:
    - bellis5
  timeout: 40

The steps Section

This section contains two subsections: build and run.

The build Subsection

Contains commands to set up the environment and compile your benchmarks. The subsection expects a multiline string with shell commands executed sequentially.

steps:
  build: |
    sudo apt-get update
    sudo apt-get install -y libblas-dev
    docc -O2 benchmark.c -o benchmark.out

The run Subsection

Defines individual benchmarks, including commands and measurement configurations: Each benchmark entry supports the following parameters:

Parameter	Type	Description
command	String	Command to execute the benchmark.
measurements	Integer	Number of repeated runs to ensure statistical reliability.
metrics	List of strings	Performance metrics collected during benchmark execution. Available metrics depend on the selected partition; see the Partitions Reference.
profiler (optional)	String	Profiling tool used to collect to performance metrics. Currently supports perf, ncu, and pyspy.
threshold (optional)	Integer	Absolute runtime threshold in seconds. Jobs exceeding this threshold are marked as failed.
loops (optional)	Boolean	Enabling the collection of profiling data for loop nests. Constructs precise roofline models. See AutomaticLoop-Level Analysis with DOCC. Only available for compilation with the DOCC.
kernels (optional)	Boolean	Enabling the collection of profiling data for individual computing kernels. Can be visualized in an interactive flame graph. Only available on GPU.

steps:
  run:
    benchmark_1:
      command: ./benchmark.out --size 1000
      measurements: 5
      threshold: 0.05
      profiler: perf
      loops: true
      kernels: false
      metrics:
        - flop_dp
        - memory_volume