benchmarking

Documentation

Index

• func ExpectFuncToComplete(f func(), runtimeThresholdInSeconds float64)
• func TimeForFuncToComplete(f func()) float64
• type Result
  • func Measure(f func()) (Result, error)

Functions

func ExpectFuncToComplete

func ExpectFuncToComplete(f func(), runtimeThresholdInSeconds float64)

ONLY COMPILES ON LINUX most often you will want to wrap the code being tested (i.e., f()) in a for loop to get more sizeable readings. use this function over built-in benchmarker (https://onsi.github.io/ginkgo/#benchmark_tests) to measure time using the OS user-time specific to the executing thread. This _should_ help with noisy neighbors in ci/locally.

other things to note while benchmarking:

• I've noticed the google cloudbuild hardware matters a lot, using nonstandard hardware (e.g. `N1_HIGHCPU_8`) has helped improve reliability of benchmarks
• we could still explore running tests with nice and/or ionice
• could also further explore running the tests with docker flags --cpu-shares set

Deprecated: use Measure instead.

func TimeForFuncToComplete

func TimeForFuncToComplete(f func()) float64

TimeForFuncToComplete returns the time the given function spend executing in user mode. Deprecated: use Measure instead.

Types

type Result

type Result struct {
	// Time spent in user mode
	Utime time.Duration
	// Time spent in kernel mode
	Stime time.Duration
	// Time spent in user mode + kernel mode
	Total time.Duration
}

Result represents the result of measuring a function's execution time.

func Measure

func Measure(f func()) (Result, error)

Measure returns the time it took to execute the given function. It only compiles on Linux. Most often you will want to run the code you want to test in a loop to get more sizeable readings, e.g.:

results := benchmarking.Measure(func() {
	for i := 0; i < 100; i++ {
		funcToTest()
	}
})

Measure should be preferred over the Gomega benchmark utils (https://pkg.go.dev/github.com/onsi/gomega/gmeasure) as it takes some additional steps to ensure we get accurate measurements.

Further ideas for improvement:

• we could explore running tests with nice and/or ionice
• could also further explore running the tests with docker flags --cpu-shares set
• consider setting GOMAXPROCS when running the tests to ensure we run in a single thread