Using Rally on Heat gates

Heat gate allows to use Rally for performance testing for each particular patch. This functionality can be used for checking patch on performance regressions and also for detecting any floating bugs for common scenarios.

How to run Rally for particular patch

As was mentioned above Heat allows to execute Rally scenarios as a gate job for particular patch. It can be done by posting comment with text check experimental for patch on review. It will run bunch of jobs, one of which has name gate-rally-dsvm-fakevirt-heat.

List of scenarios, which will be executed, is presented in file heat-fakevirt.yaml. Default version of this file is available here: https://github.com/openstack/heat/blob/master/rally-scenarios/heat-fakevirt.yaml

Obviously performance analysis make sense, when it can be compared with some another performance data. So two different approaches can be used for it:

Examples of using Rally

Previously two main approaches of using Rally job for Heat were highlighted. Corresponding examples will be described in this part of documentation.

However need to note, that there are a lot of other ways how to use Rally job for Heat performance. For example, this job can be launched periodically (twice in week) for random patches and these results will be compared between each other. It allows to see, that Heat has not any performance regressions.

Check performance or how to detect regression

The easiest way of using Rally is to execute already existing scenarios. One of the examples is presented in patch https://review.openstack.org/#/c/279450/ . In this patch was executed scenario already existing in Rally HeatStacks.create_and_delete_stack. During executing this scenario Rally creates and then, when stack is created, delete Heat stack. All existing scenarios can be found here: https://github.com/openstack/rally-openstack/blob/master/rally_openstack/scenarios/heat/stacks.py

Mentioned scenario uses Heat template as a parameter for task. The template path should be mentioned for argument template_path. It can be one of Heat templates presented in Rally repository (https://github.com/openstack/rally-openstack/tree/master/samples/tasks/scenarios/heat/templates) or new one, like it was done for mentioned patch. New added template should be placed in rally-scenarios/extra/ directory.

Also it’s possible to specify other fields for each Rally task, like sla or context. More information about other configuration setting is available by link https://rally.readthedocs.io/en/latest/plugins/#rally-plugins Mentioned patch was proposed for confirmation caching mechanism of Heat template validation process (see https://specs.openstack.org/openstack/heat-specs/specs/liberty/constraint-validation-cache.html). So it contains some changes in OS::Heat::TestResource resource, which allows to demonstrate mentioned caching feature improvements.

Initially test was run against current devstack installation, where caching is disabled (e.g. Patch Set 7). The follow results were gotten:

Action

Min (sec)

Max (sec)

Avg (sec)

Success

Count

heat.create_stack

38.223

48.085

42.971

100.0%

10

heat.delete_stack

11.755

18.155

14.085

100.0%

10

total

50.188

65.361

57.057

100.0%

10

In the next patch set (Patch Set 8) was updated by adding Depends-On reference to commit message. It let to execute the same test with patch for devstack, which turns on caching (https://review.openstack.org/#/c/279400/). The results for this case were:

Action

Min (sec)

Max (sec)

Avg (sec)

Success

Count

heat.create_stack

11.863

16.074

14.174

100.0%

10

heat.delete_stack

9.144

11.663

10.595

100.0%

10

total

21.557

27.18

24.77

100.0%

10

Comparison average values for create_stack action in the first and the second executions shows, that with enabled caching create_stack works faster in 3 times. It is a tangible improvement for create_stack operation. Need to note, that in described test delay for each constraint validation request takes 0.3 sec. as specified in constraint_prop_secs property of TestResource. It may be more, than real time delay, but it allows to confirm, that caching works correct.

Also this approach may be used for detecting regressions. In this case workflow may be presented as follow list of steps:

  • add to task list (heat-fakevirt.yaml) existing or new tasks.

  • wait a result of this execution.

  • upload patchset with changes (new feature) and launch the same test again.

  • compare performance results.

Compare output API performance

Another example of using Rally job is writing custom Rally scenarios in Heat repository. There is an example of this is presented on review: https://review.openstack.org/#/c/270225/

It’s similar on the first example, but requires more Rally specific coding. New tasks in heat-fakevirt.yaml use undefined in Rally repository scenarios:

  • CustomHeatBenchmark.create_stack_and_show_output_new

  • CustomHeatBenchmark.create_stack_and_show_output_old

  • CustomHeatBenchmark.create_stack_and_list_output_new

  • CustomHeatBenchmark.create_stack_and_list_output_old

All these scenarios are defined in the same patch and placed in rally-scenarios/plugins/ directory.

The aim of these scenarios and tasks is to demonstrate differences between new and old API calls. Heat client has a two commands for operating stack outputs: heat output-list and heat output-show <output-id>. Previously there are no special API calls for getting this information from server and this data was obtained from whole Heat Stack object. This was changed after implementation new API for outputs: https://specs.openstack.org/openstack/heat-specs/specs/mitaka/api-calls-for-output.html

As described in the mentioned specification outputs can be obtained via special requests to Heat API. According to this changes code in Heat client was updated to use new API, if it’s available.

The initial problem for this change was performance issue, which can be formulated as: execution command heat output-show <output-id> with old approach required resolving all outputs in Heat Stack, before getting only one output specified by user.

The same issue was and with heat output-list, which required to resolve all outputs only for providing list of output keys without resolved values.

Two scenarios with suffix *_new use new output API. These scenarios are not presented in Rally yet, because it’s new API. Another two scenarios with suffix *_old are based on the old approach of getting outputs. This code was partially replaced by new API, so it’s not possible to use it on fresh devstack. As result this custom code was written as two custom scenarios.

All these scenarios were added to task list and executed in the same time. Results of execution are shown below:

create_stack_and_show_output_old

Action

Min (sec)

Max (sec)

Avg (sec)

Success

Count

heat.create_stack

13.559

14.298

13.899

100.0%

5

heat.show_output_old

5.214

5.297

5.252

100.0%

5

heat.delete_stack

5.445

6.962

6.008

100.0%

5

total

24.243

26.146

25.159

100.0%

5

create_stack_and_show_output_new

Action

Min (sec)

Max (sec)

Avg (sec)

Success

Count

heat.create_stack

13.719

14.286

13.935

100.0%

5

heat.show_output_new

0.699

0.835

0.762

100.0%

5

heat.delete_stack

5.398

6.457

5.636

100.0%

5

total

19.873

21.21

20.334

100.0%

5

Average value for execution output-show for old approach obviously more, then for new API. It happens, because new API resolve only one specified output.

Same results are for output-list:

create_stack_and_list_output_old

Action

Min (sec)

Max (sec)

Avg (sec)

Success

Count

heat.create_stack

13.861

14.573

14.141

100.0%

5

heat.list_output_old

5.247

5.339

5.281

100.0%

5

heat.delete_stack

6.727

6.845

6.776

100.0%

5

total

25.886

26.696

26.199

100.0%

5

create_stack_and_list_output_new

Action

Min (sec)

Max (sec)

Avg (sec)

Success

Count

heat.create_stack

13.902

21.117

16.729

100.0%

5

heat.list_output_new

0.147

0.363

0.213

100.0%

5

heat.delete_stack

6.616

8.202

7.022

100.0%

5

total

20.838

27.908

23.964

100.0%

5

It’s also expected, because for getting list of output names is not necessary resolved values, how it is done in new API.

All mentioned results clearly show performance changes and allow to confirm, that new approach works correctly.