Analyzing Instances
Workflow execution instances have been widely used to profile and characterize workflow executions, and to build distributions of workflow execution behaviors, which are used to evaluate methods and techniques in simulation or in real conditions.
The WfCommons project targets the analysis of actual workflow execution instances (i.e., the workflow execution profile data and characterizations) in order to build Workflow Recipes of workflow applications. These recipes contain the necessary information for generating synthetic, yet realistic, workflow instances that resemble the structure and distribution of the original workflow executions.
A list of workflow execution instances that are compatible with WfFormat is kept constantly updated in our project website.
WfInstances
A workflow execution instance represents an actual execution of a scientific
workflow on a distributed platform (e.g., clouds, grids, HPC, etc.). In the
WfCommons project, an instance is represented in a JSON file following the
schema described in WfFormat. This Python package
provides an instance loader tool for importing workflow execution instances
for analysis. For instance, the code snippet below shows how an instance can
be loaded using the Instance
class:
import pathlib
from wfcommons import Instance
input_instance = pathlib.Path('/path/to/instance/file.json')
instance = Instance(input_instance=input_instance)
The Instance
class provides a number of
methods for interacting with the workflow instance, including:
draw()
: produces an image or a pdf file representing the instance.leaves()
: gets the leaves of the workflow (i.e., the tasks without any successors).roots()
: gets the roots of the workflow (i.e., the tasks without any predecessors).write_dot()
: writes a dot file of the instance.
Note
Although the analysis methods are inherently used by WfCommons (specifically WfChef) for Generating Workflows Recipes, they can also be used in a standalone manner.
The Instance Analyzer
The InstanceAnalyzer
class provides
a number of tools for analyzing collection of workflow execution instances. The
goal of the InstanceAnalyzer
is to
perform analyzes of one or multiple workflow execution instances, and build
summaries of the analyzes per workflow’ task type prefix.
Warning
Although any workflow execution instance represented as a
Instance
object (i.e., compatible with
WfFormat) can be appended to the
InstanceAnalyzer
, we strongly
recommend that only instances of a single workflow application type be
appended to an analyzer object. You may though create several analyzer
objects per workflow application.
The append_instance()
method
allows you to include instances for analysis. The
build_summary()
method
processes all appended instances. The method applies probability distributions fitting
to a series of data to find the best (i.e., minimizes the mean square error)
probability distribution that represents the analyzed data. The method returns
a summary of the analysis of instances in the form of a Python dictionary object in
which keys are task prefixes (provided when invoking the method) and values
describe the best probability distribution fit for tasks’ runtime, and input and
output data file sizes. The code excerpt below shows an example of an analysis
summary showing the best fit probability distribution for runtime of the
individuals
tasks (1000Genome workflow):
"individuals": {
"runtime": {
"min": 48.846,
"max": 192.232,
"distribution": {
"name": "skewnorm",
"params": [
11115267.652937062,
-2.9628504044929433e-05,
56.03957070238482
]
}
},
...
}
Workflow analysis summaries are used by WfChef to develop Workflow Recipes, in which themselves are used to generate realistic synthetic workflow instances.
Probability distribution fits can also be plotted by using the
generate_fit_plots()
or
generate_all_fit_plots()
methods – plots will be saved as png
files.
Examples
The following example shows the analysis of a set of instances, stored in a local folder,
of a Seismology workflow. In this example, we seek for finding the best probability
distribution fitting for task prefixes of the Seismology workflow
(sG1IterDecon
, and wrapper_siftSTFByMisfit
), and generate all fit
plots (runtime, and input and output files) into the fits
folder using
seismology
as a prefix for each generated plot:
import pathlib
from wfcommons import Instance, InstanceAnalyzer
# obtaining list of instance files in the folder
INSTANCES_PATH = pathlib.Path('/path/to/some/instance/folder/')
instance_files = [f for f in INSTANCES_PATH.glob('*') if INSTANCES_PATH.joinpath(f).is_file()]
# creating the instance analyzer object
analyzer = InstanceAnalyzer()
# appending instance files to the instance analyzer
for instance_file in instance_files:
instance = Instance(input_instance=INSTANCES_PATH.joinpath(instance_file))
analyzer.append_instance(instance)
# list of workflow task name prefixes to be analyzed in each instance
workflow_tasks = ['sG1IterDecon', 'wrapper_siftSTFByMisfit']
# building the instance summary
instances_summary = analyzer.build_summary(workflow_tasks, include_raw_data=True)
# generating all fit plots (runtime, and input and output files)
analyzer.generate_all_fit_plots(outfile_prefix='fits/seismology')