Scheduler Performance Test

This package contains the scheduler performance tests, often called scheduler_perf. We use it for benchmarking the scheduler with in-tree plugins, which is visible at perf-dash. Also you can use it outside the Kubernetes repository with out-of-tree plugins by making use of RunBenchmarkPerfScheduling.

Motivation

We already have a performance testing system -- Kubemark. However, Kubemark requires setting up and bootstrapping a whole cluster, which takes a lot of time.

We want to have a standard way to reproduce scheduling latency metrics result and benchmark scheduler as simple and fast as possible. We have the following goals:

• Save time on testing
  • The test and benchmark can be run in a single box. We only set up components necessary to scheduling without booting up a cluster.
• Profiling runtime metrics to find out bottleneck
  • Write scheduler integration test but focus on performance measurement. Take advantage of go profiling tools and collect fine-grained metrics, like cpu-profiling, memory-profiling and block-profiling.
• Reproduce test result easily
  • We want to have a known place to do the performance related test for scheduler. Developers should just run one script to collect all the information they need.

Currently the test suite has the following:

• benchmark
  • make use of go test -bench and report nanosecond/op.
  • schedule b.N pods when the cluster has N nodes and P scheduled pods. Since it takes relatively long time to finish one round, b.N is small: 10 - 100.

How To Run

Benchmark tests

# In Kubernetes root path
make test-integration WHAT=./test/integration/scheduler_perf/... ETCD_LOGLEVEL=warn KUBE_TEST_VMODULE="''" KUBE_TEST_ARGS="-run=^$$ -benchtime=1ns -bench=BenchmarkPerfScheduling"

The benchmark suite runs all the tests specified under subdirectories split by topic (<topic>/performance-config.yaml). By default, it runs all workloads that have the "performance" label. In the configuration, labels can be added to a test case and/or individual workloads. Each workload also has all labels of its test case. The perf-scheduling-label-filter command line flag can be used to select workloads. It works like GitHub label filtering: the flag accepts a comma-separated list of label names. Each label may have a + or - as prefix. Labels with + or no prefix must be set for a workload for it to be run. - means that the label must not be set. For example, this runs all performance benchmarks except those that are labeled as "integration-test":

make test-integration WHAT=./test/integration/scheduler_perf/... ETCD_LOGLEVEL=warn KUBE_TEST_VMODULE="''" KUBE_TEST_ARGS="-run=^$$ -benchtime=1ns -bench=BenchmarkPerfScheduling -perf-scheduling-label-filter=performance,-integration-test"

Once the benchmark is finished, JSON files with metrics are available in the subdirectories (test/integration/scheduler_perf/config/<topic>). Look for BenchmarkPerfScheduling_benchmark_YYYY-MM-DDTHH:MM:SSZ.json. You can use -data-items-dir to generate the metrics files elsewhere.

In case you want to run a specific test in the suite, you can specify the test through -bench flag:

Also, bench time is explicitly set to 1ns (-benchtime=1ns flag) so each test is run only once. Otherwise, the golang benchmark framework will try to run a test more than once in case it ran for less than 1s.

# In Kubernetes root path
make test-integration WHAT=./test/integration/scheduler_perf/... ETCD_LOGLEVEL=warn KUBE_TEST_VMODULE="''" KUBE_TEST_ARGS="-run=^$$ -benchtime=1ns -bench=BenchmarkPerfScheduling/SchedulingBasic/5000Nodes/5000InitPods/1000PodsToSchedule"

To produce a cpu profile:

# In Kubernetes root path
make test-integration WHAT=./test/integration/scheduler_perf/... KUBE_TIMEOUT="-timeout=3600s" ETCD_LOGLEVEL=warn KUBE_TEST_VMODULE="''" KUBE_TEST_ARGS="-run=^$$ -benchtime=1ns -bench=BenchmarkPerfScheduling -cpuprofile ~/cpu-profile.out"

How to configure benchmark tests

Configuration files located under <topic>/performance-config.yaml contain a list of templates. Each template allows to set:

• node manifest
• manifests for initial and testing pod
• number of nodes, number of initial and testing pods
• templates for PVs and PVCs
• feature gates

See op data type implementation in scheduler_perf_test.go for available operations to build WorkloadTemplate.

Initial pods create a state of a cluster before the scheduler performance measurement can begin. Testing pods are then subject to performance measurement.

The configuration files under <topic>/performance-config.yaml contain a default list of templates to cover various scenarios. In case you want to add your own, you can extend the list with new templates. It's also possible to extend op data type, respectively its underlying data types to extend configuration of possible test cases.

Logging

The default verbosity is 2 (the recommended value for production). -v can be used to change this. The log format can be changed with -logging-format=text|json. The default is to write into a log file (when using the text format) or stderr (when using JSON). Together these options allow simulating different real production configurations and to compare their performance.

During interactive debugging sessions it is possible to enable per-test output via -use-testing-log.

Log output can be quite large, in particular when running the large benchmarks and when not using -use-testing-log. For benchmarks, we want to produce that log output in a realistic way (= write to disk using the normal logging backends) and only capture the output of a specific test as part of the job results when that test failed. Therefore each test redirects its own output if the ARTIFACTS env variable is set to a $ARTIFACTS/<test name>.log file and removes that file only if the test passed.

Integration tests

To run integration tests, use:

make test-integration WHAT=./test/integration/scheduler_perf/... KUBE_TEST_ARGS=-use-testing-log

Integration testing uses the same configs (<topic>/performance-config.yaml) as benchmarking. By default, workloads labeled as integration-test are executed as part of integration testing (in ci-kubernetes-integration-master job). -test-scheduling-label-filter can be used to change that. All test cases should have at least one workload labeled as integration-test.

Running the integration tests with command above will only execute the workloads labeled as short. SHORT=--short=false variable added to the command can be used to disable this filtering.

We should make each test case with short label very small, so that all tests with the label should take less than 5 min to complete. The test cases labeled as short are executed in pull-kubernetes-integration job.

Labels used by CI jobs

See the comment on ./misc/performance-config.yaml for the details.

Scheduling throughput thresholds

Thresholds are used to capture scheduler performance regressions in a periodic ci-benchmark-scheduler-perf job. Most test cases have a threshold set for the largest performance workloads. By default, these are defined for the Average statistic of the SchedulingThroughput metric. It is possible to use other metric by configuring thresholdMetricSelector per test case or workload.

How to calculate the threshold

The initial values for scheduling throughput thresholds were calculated through an analysis of historical data, specifically focusing on the minimum, average, and standard deviation values for each workload (see #126871). Our goal is to set the thresholds somewhat pessimistically to minimize flakiness, so it's recommended to set the threshold slightly below the observed historical minimum. Depending on variability of data, the threshold can be lowered more.

Thresholds should be adjusted based on the flakiness level and minima observed in the future. Remember to set the value for newly added test cases as well, but after collecting some data on workload characteristics.

How to determine the failed workload

When the workload's scheduling throughput doesn't exceed the threshold, the ci-benchmark-scheduler-perf periodic job will fail with an error log such as:

--- FAIL: BenchmarkPerfScheduling/SchedulingBasic/5000Nodes_10000Pods
    ...
    scheduler_perf.go:1098: ERROR: op 2: SchedulingBasic/5000Nodes_10000Pods/namespace-2: expected SchedulingThroughput Average to be higher: got 256.12, want 270

This allows to analyze which workload failed. Make sure that the failure is not an outlier by checking multiple runs of the job. If the failures are not related to any regression, but to an incorrect threshold setting, it is reasonable to decrease it.

Visualization

Some support for visualizing progress over time is built into the benchmarks. The measurement operation which creates pods writes .dat files like this:

 test/integration/scheduler_perf/misc/SchedulingBasic_5000Nodes_2023-03-17T14:52:09Z.dat

This file is in a text format that gnuplot can read. A wrapper script selects some suitable parameters:

 test/integration/scheduler_perf/gnuplot.sh test/integration/scheduler_perf/*/*.dat

It plots in an interactive window by default. To write into a file, use

test/integration/scheduler_perf/gnuplot.sh \
   -e 'set term png; set output "<output>.png"' \
   test/integration/scheduler_perf/*/*.dat