metrics

Kata Containers metrics

This directory contains the metrics tests for Kata Containers.

The tests within this directory have a number of potential use cases:

• CI checks for regressions on PRs
• CI data gathering for main branch merges
• Developer use for pre-checking code changes before raising a PR
• As part of report generation

Goals

This section details some of the goals for the potential use cases.

PR regression checks

The goal for the PR CI regression checking is to provide a relatively quick CI metrics check and feedback directly back to the GitHub PR.

Due to the relatively fast feedback requirement, there is generally a compromise that has to be made with the metrics - precision vs time.

Therefore, there is a separate script which invokes a subset of the metrics tests, configured for speed over accuracy.

Having said that, accuracy is still important. If we have very noisy tests, then the CI will either not spot regressions that are below that noise factor, or will cause false failures, which are very undesirable in a CI.

Developer pre-checking

The PR regression check scripts can be executed "by hand", and thus are available for developers to use as a "pre-check" before submitting a PR. It might be prudent for developers to follow this procedure particularly for large architectural or version changes of components.

Stability or Performance?

When forming, or configuring, metrics tests, often we have to make a choice or compromise on if we want the test to take a more repeatable measurement (less noise and variance), or if we want to measure the "best performance".

Generally for CI regression checking, we prefer stability over performance, as that allows us a more accurate (narrow bound) check for regressions.

NOTE It should thus be noted that if you are gathering data to discuss performance, then you may not want to use the CI metrics data, as this may be favoring stable results over best performance.

Requirements

To try and maintain the quality of the metrics data gathered and the accuracy of the CI regression checking, we try to define and stick to some "quality measures" for our metrics.

For PR checks

The PR CI is generally required to execute within a "reasonable time" to provide timely feedback to the developers (and not stall the review and development process). To that end, we relax the quality requirements of the PR CI. The quality requirements are:

• <= 5% run to run variance
• <= 5 minutes runtime per test

Categories

Kata Container metrics tend to fall into a set of categories, and we organise the tests within this folder as such.

Each sub-folder contains its own README detailing its own tests.

Time (Speed)

Generally tests that measure the "speed" of the runtime itself, such as time to boot into a workload or kill a container.

This directory does not contain "speed" tests that measure network or storage for instance.

For further details see the time tests documentation.

Density

Tests that measure the size and overheads of the runtime. Generally this is looking at memory footprint sizes, but could also cover disk space or even CPU consumption.

For further details see the density tests documentation.

Networking

Tests relating to networking. General items could include:

• bandwidth
• latency
• jitter
• parallel bandwidth
• write and read percentiles

For further details see the network tests documentation.

Storage

Tests relating to the storage (graph, volume) drivers. Measures may include:

• bandwidth
• latency
• jitter
• conformance (to any relevant standards)

For further details see the storage tests documentation.

Disk

Test relating to measure reading and writing against clusters.

For further details see the disk tests documentation.

Machine Learning

Tests relating with TensorFlow implementations of several popular convolutional models.

For further details see the machine learning tests documentation.

Saving Results

In order to ensure continuity, and thus testing and historical tracking of results, we provide a bash API to aid storing results in a uniform manner.

JSON API

The preferred API to store results is through the provided JSON API.

The API provides the following groups of functions:

• A set of functions to init/save the data and add "top level" JSON fragments
• A set of functions to construct arrays of JSON fragments, which are then added as a top level fragment when complete
• A set of functions to construct elements of an array from sub-fragments, and then finalize that element when all fragments are added.

Construction of JSON data under bash could be relatively complex. This API does not pretend to support all possible data constructs or features, and individual tests may find they need to do some JSON handling themselves before injecting their JSON into the API.

If you find a common use case that many tests are implementing themselves, then please factor out that functionality and consider extending this API.

metrics_json_init()

Initialise the API. Must be called before all other JSON API calls. Should be matched by a final call to metrics_json_save.

Relies upon the TEST_NAME variable to derive the file name the final JSON data is stored in (under the metrics/results directory). If your test generates multiple TEST_NAME sets of data then:

• Ensure you have a matching JSON init/save call pair for each of those sets.
• These sets could be a hangover from a previous CSV based test - consider using a single JSON file if possible to store all the results.

This function may add system level information to the results file as a top level fragment, for example:

• env - A fragment containing system level environment information
• "time" - A fragment containing a nanosecond timestamp of when the test was executed

Consider these top level JSON section names to be reserved by the API.

metrics_json_save()

This function saves all registered JSON fragments out to the JSON results file.

Note: this function will not save any part-registered array fragments. They will be lost.

metrics_json_add_fragment(json)

Add a JSON formatted fragment at the top level.

Arg	Description
json	A fully formed JSON fragment

metrics_json_start_array()

Initialise the JSON array API subsystem, ready to accept JSON fragments via metrics_json_add_array_element.

This JSON array API subset allows accumulation of multiple entries into a JSON [] array, to later be added as a top level fragment.

metrics_json_add_array_element(json)

Add a fully formed JSON fragment to the JSON array store.

Arg	Description
json	A fully formed JSON fragment

metrics_json_add_array_fragment(json)

Add a fully formed JSON fragment to the current array element.

Arg	Description
json	A fully formed JSON fragment

metrics_json_close_array_element()

Finalize (close) the current array element. This incorporates any array_fragment parts into the current array element, closes that array element, and reset the in-flight array_fragment store.

metrics_json_end_array(name)

Save the stored JSON array store as a top level fragment, with the name name.

Arg	Description
name	The name to be given to the generated top level fragment array

Preserving results

The JSON library contains a hook that enables results to be injected to a data store at the same time they are saved to the results files.

The hook supports transmission via curl or socat. Configuration is via environment variables.

Variable	Description
JSON_HOST	Destination host path for use with socat
JSON_SOCKET	Destination socket number for use with socat
JSON_URL	Destination URL for use with curl
JSON_TX_ONELINE	If set, the JSON will be sent as a single line (CR and tabs stripped)

socat transmission will only happen if JSON_HOST is set. curl transmission will only happen if JSON_URL is set. The settings are not mutually exclusive, and both can be set if necessary.

JSON_TX_ONELINE applies to both types of transmission.

Report generator

See the report generator documentation.