From 0a4f0e9f32489ad8a3a15448d19349755830da84 Mon Sep 17 00:00:00 2001 From: alitto Date: Sat, 6 Jun 2020 16:37:53 -0300 Subject: [PATCH] Simplify resizing strategies --- README.md | 66 ++++--- benchmark/benchmark_test.go | 344 +++++++++++++++++++--------------- docs/strategies.svg | 1 + pond.go | 360 ++++++++++-------------------------- pond_blackbox_test.go | 49 +++++ resizer.go | 188 +++---------------- resizer_test.go | 50 ++--- 7 files changed, 430 insertions(+), 628 deletions(-) create mode 100644 docs/strategies.svg diff --git a/README.md b/README.md index ced26fa..b296457 100644 --- a/README.md +++ b/README.md @@ -158,16 +158,24 @@ panicHandler := func(p interface{}) { pool := pond.New(10, 1000, pond.PanicHandler(panicHandler))) ``` - **Strategy**: Configures the strategy used to resize the pool when backpressure is detected. You can create a custom strategy by implementing the `pond.ResizingStrategy` interface or choose one of the 3 presets: - - **Eager**: maximizes responsiveness at the expense of higher resource usage, which can reduce throughput under certain conditions. This strategy is meant for worker pools that will operate at a small percentage of their capacity most of the time and may occasionally receive bursts of tasks. - - **Balanced**: tries to find a balance between responsiveness and throughput. It's suitable for general purpose worker pools or those that will operate close to 50% of their capacity most of the time. This is the default strategy. + - **Eager**: maximizes responsiveness at the expense of higher resource usage, which can reduce throughput under certain conditions. This strategy is meant for worker pools that will operate at a small percentage of their capacity most of the time and may occasionally receive bursts of tasks. This is the default strategy. + - **Balanced**: tries to find a balance between responsiveness and throughput. It's suitable for general purpose worker pools or those that will operate close to 50% of their capacity most of the time. - **Lazy**: maximizes throughput at the expense of responsiveness. This strategy is meant for worker pools that will operate close to their max. capacity most of the time. ``` go // Example: create pools with different resizing strategies -eagerPool := pond.New(10, 1000, pond.Strategy(pond.Eager)) -balancedPool := pond.New(10, 1000, pond.Strategy(pond.Balanced)) -lazyPool := pond.New(10, 1000, pond.Strategy(pond.Lazy)) +eagerPool := pond.New(10, 1000, pond.Strategy(pond.Eager())) +balancedPool := pond.New(10, 1000, pond.Strategy(pond.Balanced())) +lazyPool := pond.New(10, 1000, pond.Strategy(pond.Lazy())) ``` +### Resizing strategies + +The following chart illustrates the behaviour of the different pool resizing strategies as the number of submitted tasks increases. Each line represents the number of worker goroutines in the pool (pool size) and the x-axis reflects the number of submitted tasks (cumulative). + +![Pool resizing strategies behaviour](./docs/strategies.svg) + +As the name suggests, the "Eager" strategy always spawns an extra worker when there are no idles, which causes the pool to grow almost linearly with the number of submitted tasks. On the other end, the "Lazy" strategy creates one worker every N submitted tasks, where N is the maximum number of available CPUs ([GOMAXPROCS](https://golang.org/pkg/runtime/#GOMAXPROCS)). The "Balanced" strategy represents a middle ground between the previous two because it creates a worker every N/2 submitted tasks. + ## API Reference Full API reference is available at https://pkg.go.dev/github.com/alitto/pond @@ -189,30 +197,30 @@ Here are the results: goos: linux goarch: amd64 pkg: github.com/alitto/pond/benchmark -BenchmarkAll/1M-10ms/Pond-Eager-8 2 620347142 ns/op 82768720 B/op 1086686 allocs/op -BenchmarkAll/1M-10ms/Pond-Balanced-8 2 578973910 ns/op 81339088 B/op 1083203 allocs/op -BenchmarkAll/1M-10ms/Pond-Lazy-8 2 613344573 ns/op 84347248 B/op 1084987 allocs/op -BenchmarkAll/1M-10ms/Goroutines-8 2 540765682 ns/op 98457168 B/op 1060433 allocs/op -BenchmarkAll/1M-10ms/GoroutinePool-8 1 1157705614 ns/op 68137088 B/op 1409763 allocs/op -BenchmarkAll/1M-10ms/BufferedPool-8 1 1158068370 ns/op 76426272 B/op 1412739 allocs/op -BenchmarkAll/1M-10ms/Gammazero-8 1 1330312458 ns/op 34524328 B/op 1029692 allocs/op -BenchmarkAll/1M-10ms/AntsPool-8 2 724231628 ns/op 37870404 B/op 1077297 allocs/op -BenchmarkAll/100k-500ms/Pond-Eager-8 2 604180003 ns/op 31523028 B/op 349877 allocs/op -BenchmarkAll/100k-500ms/Pond-Balanced-8 1 1060079592 ns/op 35520416 B/op 398779 allocs/op -BenchmarkAll/100k-500ms/Pond-Lazy-8 1 1053705909 ns/op 35040512 B/op 392696 allocs/op -BenchmarkAll/100k-500ms/Goroutines-8 2 551869174 ns/op 8000016 B/op 100001 allocs/op -BenchmarkAll/100k-500ms/GoroutinePool-8 2 635442074 ns/op 20764560 B/op 299632 allocs/op -BenchmarkAll/100k-500ms/BufferedPool-8 2 641683384 ns/op 21647840 B/op 299661 allocs/op -BenchmarkAll/100k-500ms/Gammazero-8 2 667449574 ns/op 16241864 B/op 249664 allocs/op -BenchmarkAll/100k-500ms/AntsPool-8 2 659853037 ns/op 37300372 B/op 549784 allocs/op -BenchmarkAll/10k-1000ms/Pond-Eager-8 1 1014320653 ns/op 12135080 B/op 39692 allocs/op -BenchmarkAll/10k-1000ms/Pond-Balanced-8 1 1015979207 ns/op 12083704 B/op 39518 allocs/op -BenchmarkAll/10k-1000ms/Pond-Lazy-8 1 1036374161 ns/op 12046632 B/op 39366 allocs/op -BenchmarkAll/10k-1000ms/Goroutines-8 1 1007837894 ns/op 800016 B/op 10001 allocs/op -BenchmarkAll/10k-1000ms/GoroutinePool-8 1 1149536612 ns/op 21393024 B/op 222458 allocs/op -BenchmarkAll/10k-1000ms/BufferedPool-8 1 1127286218 ns/op 20343584 B/op 219359 allocs/op -BenchmarkAll/10k-1000ms/Gammazero-8 1 1023249222 ns/op 2019688 B/op 29374 allocs/op -BenchmarkAll/10k-1000ms/AntsPool-8 1 1016280850 ns/op 4155904 B/op 59487 allocs/op +1M-10ms/Pond-Eager-8 2 620347142 82768720 1086686 +1M-10ms/Pond-Balanced-8 2 578973910 81339088 1083203 +1M-10ms/Pond-Lazy-8 2 613344573 84347248 1084987 +1M-10ms/Goroutines-8 2 540765682 98457168 1060433 +1M-10ms/GoroutinePool-8 1 1157705614 68137088 1409763 +1M-10ms/BufferedPool-8 1 1158068370 76426272 1412739 +1M-10ms/Gammazero-8 1 1330312458 34524328 1029692 +1M-10ms/AntsPool-8 2 724231628 37870404 1077297 +100k-500ms/Pond-Eager-8 2 604180003 31523028 349877 +100k-500ms/Pond-Balanced-8 1 1060079592 35520416 398779 +100k-500ms/Pond-Lazy-8 1 1053705909 35040512 392696 +100k-500ms/Goroutines-8 2 551869174 8000016 100001 +100k-500ms/GoroutinePool-8 2 635442074 20764560 299632 +100k-500ms/BufferedPool-8 2 641683384 21647840 299661 +100k-500ms/Gammazero-8 2 667449574 16241864 249664 +100k-500ms/AntsPool-8 2 659853037 37300372 549784 +10k-1000ms/Pond-Eager-8 1 1014320653 12135080 39692 +10k-1000ms/Pond-Balanced-8 1 1015979207 12083704 39518 +10k-1000ms/Pond-Lazy-8 1 1036374161 12046632 39366 +10k-1000ms/Goroutines-8 1 1007837894 800016 10001 +10k-1000ms/GoroutinePool-8 1 1149536612 21393024 222458 +10k-1000ms/BufferedPool-8 1 1127286218 20343584 219359 +10k-1000ms/Gammazero-8 1 1023249222 2019688 29374 +10k-1000ms/AntsPool-8 1 1016280850 4155904 59487 PASS ok github.com/alitto/pond/benchmark 37.331s ``` diff --git a/benchmark/benchmark_test.go b/benchmark/benchmark_test.go index 9ff1664..d52881c 100644 --- a/benchmark/benchmark_test.go +++ b/benchmark/benchmark_test.go @@ -2,6 +2,7 @@ package benchmark import ( "fmt" + "math/rand" "sync" "testing" "time" @@ -11,193 +12,236 @@ import ( "github.com/panjf2000/ants/v2" ) +type subject struct { + name string + factory poolFactory +} + +type poolSubmit func(func()) +type poolTeardown func() +type poolFactory func() (poolSubmit, poolTeardown) + type workload struct { name string + userCount int taskCount int - taskDuration time.Duration -} + taskInterval time.Duration + task func() +} + +var maxWorkers = 200000 + +var workloads = []workload{{ + name: "1u-10Mt", + userCount: 1, + taskCount: 1000000, + taskInterval: 0, +}, { + name: "100u-10Kt", + userCount: 100, + taskCount: 10000, + taskInterval: 0, +}, { + name: "1Ku-1Kt", + userCount: 1000, + taskCount: 1000, + taskInterval: 0, +}, { + name: "10Ku-100t", + userCount: 10000, + taskCount: 100, + taskInterval: 0, +}, { + name: "1Mu-1t", + userCount: 1000000, + taskCount: 1, + taskInterval: 0, +}} -type subject struct { - name string - test poolTest - config poolConfig +var pondSubjects = []subject{ + { + name: "Pond-Eager", + factory: func() (poolSubmit, poolTeardown) { + pool := pond.New(maxWorkers, 1000000, pond.Strategy(pond.Eager())) + + return pool.Submit, pool.StopAndWait + }, + }, { + name: "Pond-Balanced", + factory: func() (poolSubmit, poolTeardown) { + pool := pond.New(maxWorkers, 1000000, pond.Strategy(pond.Balanced())) + + return pool.Submit, pool.StopAndWait + }, + }, { + name: "Pond-Lazy", + factory: func() (poolSubmit, poolTeardown) { + pool := pond.New(maxWorkers, 1000000, pond.Strategy(pond.Lazy())) + + return pool.Submit, pool.StopAndWait + }, + }, } -type poolConfig struct { - minWorkers int - maxWorkers int - maxCapacity int - strategy pond.ResizingStrategy -} +var otherSubjects = []subject{ + { + name: "Goroutines", + factory: func() (poolSubmit, poolTeardown) { + submit := func(taskFunc func()) { + go func() { + taskFunc() + }() + } + return submit, func() {} + }, + }, + { + name: "GoroutinePool", + factory: func() (poolSubmit, poolTeardown) { + + var poolWg sync.WaitGroup + taskChan := make(chan func()) + poolWg.Add(maxWorkers) + for i := 0; i < maxWorkers; i++ { + go func() { + for task := range taskChan { + task() + } + poolWg.Done() + }() + } + + submit := func(task func()) { + taskChan <- task + } + teardown := func() { + close(taskChan) + poolWg.Wait() + } -type poolTest func(taskCount int, taskFunc func(), config poolConfig) + return submit, teardown + }, + }, + { + name: "BufferedPool", + factory: func() (poolSubmit, poolTeardown) { + + var poolWg sync.WaitGroup + taskChan := make(chan func(), 1000000) + poolWg.Add(maxWorkers) + for i := 0; i < maxWorkers; i++ { + go func() { + for task := range taskChan { + task() + } + poolWg.Done() + }() + } -var workloads = []workload{ - {"1M-10ms", 1000000, 10 * time.Millisecond}, - {"100k-500ms", 100000, 500 * time.Millisecond}, - {"10k-1000ms", 10000, 1000 * time.Millisecond}, -} + submit := func(task func()) { + taskChan <- task + } + teardown := func() { + close(taskChan) + poolWg.Wait() + } -var defaultPoolConfig = poolConfig{ - maxWorkers: 200000, + return submit, teardown + }, + }, + { + name: "Gammazero", + factory: func() (poolSubmit, poolTeardown) { + pool := workerpool.New(maxWorkers) + return pool.Submit, pool.StopWait + }, + }, + { + name: "AntsPool", + factory: func() (poolSubmit, poolTeardown) { + pool, _ := ants.NewPool(maxWorkers, ants.WithExpiryDuration(10*time.Second)) + submit := func(task func()) { + pool.Submit(task) + } + return submit, pool.Release + }, + }, } -var pondSubjects = []subject{ - {"Pond-Eager", pondPool, poolConfig{maxWorkers: defaultPoolConfig.maxWorkers, maxCapacity: 1000000, strategy: pond.Eager()}}, - {"Pond-Balanced", pondPool, poolConfig{maxWorkers: defaultPoolConfig.maxWorkers, maxCapacity: 1000000, strategy: pond.Balanced()}}, - {"Pond-Lazy", pondPool, poolConfig{maxWorkers: defaultPoolConfig.maxWorkers, maxCapacity: 1000000, strategy: pond.Lazy()}}, +func BenchmarkPondSleep10ms(b *testing.B) { + sleep10ms := func() { + time.Sleep(10 * time.Millisecond) + } + runBenchmarks(b, workloads, pondSubjects, sleep10ms) } -var otherSubjects = []subject{ - {"Goroutines", unboundedGoroutines, defaultPoolConfig}, - {"GoroutinePool", goroutinePool, defaultPoolConfig}, - {"BufferedPool", bufferedGoroutinePool, defaultPoolConfig}, - {"Gammazero", gammazeroWorkerpool, defaultPoolConfig}, - {"AntsPool", antsPool, defaultPoolConfig}, +func BenchmarkPondRandFloat64(b *testing.B) { + randFloat64 := func() { + rand.Float64() + } + runBenchmarks(b, workloads, pondSubjects, randFloat64) } -func BenchmarkPond(b *testing.B) { - runBenchmarks(b, workloads, pondSubjects) +func BenchmarkAllSleep10ms(b *testing.B) { + subjects := make([]subject, 0) + subjects = append(subjects, pondSubjects...) + subjects = append(subjects, otherSubjects...) + sleep10ms := func() { + time.Sleep(10 * time.Millisecond) + } + runBenchmarks(b, workloads, subjects, sleep10ms) } -func BenchmarkAll(b *testing.B) { - allSubjects := make([]subject, 0) - allSubjects = append(allSubjects, pondSubjects...) - allSubjects = append(allSubjects, otherSubjects...) - runBenchmarks(b, workloads, allSubjects) +func BenchmarkAllRandFloat64(b *testing.B) { + subjects := make([]subject, 0) + subjects = append(subjects, pondSubjects...) + subjects = append(subjects, otherSubjects...) + randFloat64 := func() { + rand.Float64() + } + runBenchmarks(b, workloads, subjects, randFloat64) } -func runBenchmarks(b *testing.B, workloads []workload, subjects []subject) { +func runBenchmarks(b *testing.B, workloads []workload, subjects []subject, task func()) { for _, workload := range workloads { - taskFunc := func() { - time.Sleep(workload.taskDuration) - } for _, subject := range subjects { - name := fmt.Sprintf("%s/%s", workload.name, subject.name) - b.Run(name, func(b *testing.B) { + testName := fmt.Sprintf("%s/%s", workload.name, subject.name) + b.Run(testName, func(b *testing.B) { for i := 0; i < b.N; i++ { - subject.test(workload.taskCount, taskFunc, subject.config) + simulateWorkload(&workload, subject.factory, task) } }) } } } -func pondPool(taskCount int, taskFunc func(), config poolConfig) { - var wg sync.WaitGroup - pool := pond.New(config.maxWorkers, config.maxCapacity, - pond.MinWorkers(config.minWorkers), - pond.Strategy(config.strategy)) - // Submit tasks - wg.Add(taskCount) - for n := 0; n < taskCount; n++ { - pool.Submit(func() { - taskFunc() - wg.Done() - }) - } - wg.Wait() - pool.StopAndWait() -} - -func unboundedGoroutines(taskCount int, taskFunc func(), config poolConfig) { - var wg sync.WaitGroup - wg.Add(taskCount) - for i := 0; i < taskCount; i++ { - go func() { - taskFunc() - wg.Done() - }() - } - wg.Wait() -} +func simulateWorkload(workload *workload, poolFactoy poolFactory, task func()) { -func goroutinePool(taskCount int, taskFunc func(), config poolConfig) { - // Start worker goroutines - var poolWg sync.WaitGroup - taskChan := make(chan func()) - poolWg.Add(config.maxWorkers) - for i := 0; i < config.maxWorkers; i++ { - go func() { - for task := range taskChan { - task() - } - poolWg.Done() - }() - } + // Create pool + poolSubmit, poolTeardown := poolFactoy() - // Submit tasks and wait for completion + // Spawn one goroutine per simulated user var wg sync.WaitGroup - wg.Add(taskCount) - for i := 0; i < taskCount; i++ { - taskChan <- func() { - taskFunc() - wg.Done() - } + wg.Add(workload.userCount * workload.taskCount) + + testFunc := func() { + task() + wg.Done() } - close(taskChan) - wg.Wait() - poolWg.Wait() -} -func bufferedGoroutinePool(taskCount int, taskFunc func(), config poolConfig) { - // Start worker goroutines - var poolWg sync.WaitGroup - taskChan := make(chan func(), taskCount) - poolWg.Add(config.maxWorkers) - for i := 0; i < config.maxWorkers; i++ { + for i := 0; i < workload.userCount; i++ { go func() { - for task := range taskChan { - task() + // Every user submits tasksPerUser at the specified frequency + for i := 0; i < workload.taskCount; i++ { + poolSubmit(testFunc) + if workload.taskInterval > 0 { + time.Sleep(workload.taskInterval) + } } - poolWg.Done() }() } - - // Submit tasks and wait for completion - var wg sync.WaitGroup - wg.Add(taskCount) - for i := 0; i < taskCount; i++ { - taskChan <- func() { - taskFunc() - wg.Done() - } - } - close(taskChan) - wg.Wait() - poolWg.Wait() -} - -func gammazeroWorkerpool(taskCount int, taskFunc func(), config poolConfig) { - // Create pool - wp := workerpool.New(config.maxWorkers) - defer wp.StopWait() - - // Submit tasks and wait for completion - var wg sync.WaitGroup - wg.Add(taskCount) - for i := 0; i < taskCount; i++ { - wp.Submit(func() { - taskFunc() - wg.Done() - }) - } wg.Wait() -} - -func antsPool(taskCount int, taskFunc func(), config poolConfig) { - // Create pool - pool, _ := ants.NewPool(config.maxWorkers, ants.WithExpiryDuration(10*time.Second)) - defer pool.Release() - // Submit tasks and wait for completion - var wg sync.WaitGroup - wg.Add(taskCount) - for i := 0; i < taskCount; i++ { - _ = pool.Submit(func() { - taskFunc() - wg.Done() - }) - } - wg.Wait() + // Tear down + poolTeardown() } diff --git a/docs/strategies.svg b/docs/strategies.svg new file mode 100644 index 0000000..ad919bf --- /dev/null +++ b/docs/strategies.svg @@ -0,0 +1 @@ + \ No newline at end of file diff --git a/pond.go b/pond.go index cb498c1..db4f37e 100644 --- a/pond.go +++ b/pond.go @@ -2,7 +2,6 @@ package pond import ( "fmt" - "math" "runtime/debug" "sync" "sync/atomic" @@ -22,7 +21,7 @@ func defaultPanicHandler(panic interface{}) { // ResizingStrategy represents a pool resizing strategy type ResizingStrategy interface { - Resize(runningWorkers, idleWorkers, minWorkers, maxWorkers, incomingTasks, completedTasks int, delta time.Duration) int + Resize(runningWorkers, minWorkers, maxWorkers int) bool } // Option represents an option that can be passed when instantiating a worker pool to customize it @@ -66,19 +65,13 @@ type WorkerPool struct { strategy ResizingStrategy panicHandler func(interface{}) // Atomic counters - workerCount int32 - idleWorkerCount int32 - completedTaskCount uint64 + workerCount int32 + idleWorkerCount int32 // Private properties - tasks chan func() - dispatchedTasks chan func() - stopOnce sync.Once - waitGroup sync.WaitGroup - lastResizeTime time.Time - lastResizeCompletedTasks uint64 - // Debug information - debug bool - maxWorkerCount int + tasks chan func() + purgerQuit chan struct{} + stopOnce sync.Once + waitGroup sync.WaitGroup } // New creates a worker pool with that can scale up to the given maximum number of workers (maxWorkers). @@ -92,9 +85,8 @@ func New(maxWorkers, maxCapacity int, options ...Option) *WorkerPool { maxWorkers: maxWorkers, maxCapacity: maxCapacity, idleTimeout: defaultIdleTimeout, - strategy: Balanced(), + strategy: Eager(), panicHandler: defaultPanicHandler, - debug: false, } // Apply all options @@ -118,19 +110,21 @@ func New(maxWorkers, maxCapacity int, options ...Option) *WorkerPool { // Create internal channels pool.tasks = make(chan func(), pool.maxCapacity) - pool.dispatchedTasks = make(chan func(), pool.maxWorkers) + pool.purgerQuit = make(chan struct{}) - // Start dispatcher goroutine + // Start purger goroutine pool.waitGroup.Add(1) go func() { defer pool.waitGroup.Done() - pool.dispatch() + pool.purge() }() // Start minWorkers workers if pool.minWorkers > 0 { - pool.startWorkers(pool.minWorkers, nil) + for i := 0; i < pool.minWorkers; i++ { + pool.startWorker(nil) + } } return pool @@ -147,14 +141,51 @@ func (p *WorkerPool) Idle() int { } // Submit sends a task to this worker pool for execution. If the queue is full, -// it will wait until the task can be enqueued +// it will wait until the task is dispatched to a worker goroutine. func (p *WorkerPool) Submit(task func()) { + p.submit(task, true) +} + +// TrySubmit attempts to send a task to this worker pool for execution. If the queue is full, +// it will not wait for a worker to become idle. It returns true if it was able to dispatch +// the task and false otherwise. +func (p *WorkerPool) TrySubmit(task func()) bool { + return p.submit(task, false) +} + +func (p *WorkerPool) submit(task func(), waitForIdle bool) bool { if task == nil { - return + return false + } + + runningWorkerCount := p.Running() + + // Attempt to dispatch to an idle worker without blocking + if runningWorkerCount > 0 && p.Idle() > 0 { + select { + case p.tasks <- task: + return true + default: + // No idle worker available, continue + } + } + + maxWorkersReached := runningWorkerCount >= p.maxWorkers + + // Exit if we have reached the max. number of workers and can't wait for an idle worker + if maxWorkersReached && !waitForIdle { + return false + } + + // Start a worker as long as we haven't reached the limit + if !maxWorkersReached && p.strategy.Resize(runningWorkerCount, p.minWorkers, p.maxWorkers) { + p.startWorker(task) + return true } - // Submit the task to the task channel + // Submit the task to the tasks channel and wait for it to be picked up by a worker p.tasks <- task + return true } // SubmitAndWait sends a task to this worker pool for execution and waits for it to complete @@ -196,8 +227,8 @@ func (p *WorkerPool) SubmitBefore(task func(), deadline time.Duration) { // Stop causes this pool to stop accepting tasks, without waiting for goroutines to exit func (p *WorkerPool) Stop() { p.stopOnce.Do(func() { - // Close the tasks channel to prevent receiving new tasks - close(p.tasks) + // Send the signal to stop the purger goroutine + close(p.purgerQuit) }) } @@ -209,245 +240,50 @@ func (p *WorkerPool) StopAndWait() { p.waitGroup.Wait() } -// dispatch represents the work done by the dispatcher goroutine -func (p *WorkerPool) dispatch() { - - // Declare vars - var ( - maxBatchSize = 1000 - batch = make([]func(), maxBatchSize) - batchSize = int(math.Max(float64(p.minWorkers), 100)) - idleWorkers = 0 - dispatchedToIdleWorkers = 0 - dispatchedToNewWorkers = 0 - dispatchedBlocking = 0 - nextTask func() = nil - ) - - idleTimer := time.NewTimer(p.idleTimeout) - defer idleTimer.Stop() - - // Start dispatching cycle -DispatchCycle: - for { - // Reset idle timer - idleTimer.Reset(p.idleTimeout) - - select { - // Receive a task - case task, ok := <-p.tasks: - if !ok { - // Received the signal to exit - break DispatchCycle - } - - idleWorkers = p.Idle() - - // Dispatch tasks to idle workers - nextTask, dispatchedToIdleWorkers = p.dispatchToIdleWorkers(task, idleWorkers) - if nextTask == nil { - continue DispatchCycle - } - - // Read up to batchSize tasks without blocking - p.receiveBatch(nextTask, &batch, batchSize) - - // Resize the pool - dispatchedToNewWorkers = p.resizePool(batch, dispatchedToIdleWorkers) - - dispatchedBlocking = 0 - if len(batch) > dispatchedToNewWorkers { - for _, task := range batch[dispatchedToNewWorkers:] { - // Attempt to dispatch the task without blocking - select { - case p.dispatchedTasks <- task: - default: - // Block until a worker accepts this task - p.dispatchedTasks <- task - dispatchedBlocking++ - } - } - } - - // Adjust batch size - if dispatchedBlocking > 0 { - if batchSize > 1 { - batchSize = 1 - } - } else { - batchSize = batchSize * 2 - if batchSize > maxBatchSize { - batchSize = maxBatchSize - } - } - // Timed out waiting for any activity to happen, attempt to resize the pool - case <-idleTimer.C: - p.resizePool(batch[:0], 0) - } - } - - // Send signal to stop all workers - close(p.dispatchedTasks) - - if p.debug { - fmt.Printf("Max workers: %d", p.maxWorkerCount) - } -} - -func (p *WorkerPool) dispatchToIdleWorkers(task func(), limit int) (nextTask func(), dispatched int) { - - // Dispatch up to limit tasks without blocking - nextTask = task - for i := 0; i < limit; i++ { - - // Attempt to dispatch without blocking - select { - case p.dispatchedTasks <- nextTask: - nextTask = nil - dispatched++ - default: - // Could not dispatch, return the task - return - } - - // Attempt to receive another task - select { - case t, ok := <-p.tasks: - if !ok { - // Nothing else to dispatch - nextTask = nil - return - } - nextTask = t - default: - nextTask = nil - return - } - } - - return -} - -func (p *WorkerPool) receiveBatch(task func(), batch *[]func(), batchSize int) { +// purge represents the work done by the purger goroutine +func (p *WorkerPool) purge() { - // Reset batch slice - *batch = (*batch)[:0] - *batch = append(*batch, task) + idleTicker := time.NewTicker(p.idleTimeout) + defer idleTicker.Stop() - // Read up to batchSize tasks without blocking - for i := 0; i < batchSize-1; i++ { +Purge: + for { select { - case t, ok := <-p.tasks: - if !ok { - return + // Timed out waiting for any activity to happen, attempt to kill an idle worker + case <-idleTicker.C: + if p.Idle() > 0 { + p.tasks <- nil } - if t != nil { - *batch = append(*batch, t) - } - default: - return + case <-p.purgerQuit: + break Purge } } -} - -func (p *WorkerPool) resizePool(batch []func(), dispatchedToIdleWorkers int) int { - // Time to resize the pool - now := time.Now() - workload := len(batch) - currentCompletedTasks := atomic.LoadUint64(&p.completedTaskCount) - completedTasksDelta := int(currentCompletedTasks - p.lastResizeCompletedTasks) - if completedTasksDelta < 0 { - completedTasksDelta = 0 - } - duration := 0 * time.Millisecond - if !p.lastResizeTime.IsZero() { - duration = now.Sub(p.lastResizeTime) - } - poolSizeDelta := p.calculatePoolSizeDelta(p.Running(), p.Idle(), - workload+dispatchedToIdleWorkers, completedTasksDelta, duration) - - // Capture values for next resize cycle - p.lastResizeTime = now - p.lastResizeCompletedTasks = currentCompletedTasks - - // Start up to poolSizeDelta workers - dispatched := 0 - if poolSizeDelta > 0 { - p.startWorkers(poolSizeDelta, batch) - dispatched = workload - if poolSizeDelta < workload { - dispatched = poolSizeDelta - } - } else if poolSizeDelta < 0 { - // Kill poolSizeDelta workers - for i := 0; i < -poolSizeDelta; i++ { - p.dispatchedTasks <- nil - } - } + // Send signal to stop all workers + close(p.tasks) - return dispatched } -// calculatePoolSizeDelta calculates what's the delta to reach the ideal pool size based on the current size and workload -func (p *WorkerPool) calculatePoolSizeDelta(runningWorkers, idleWorkers, - incomingTasks, completedTasks int, duration time.Duration) int { - - delta := p.strategy.Resize(runningWorkers, idleWorkers, p.minWorkers, p.maxWorkers, - incomingTasks, completedTasks, duration) - - targetSize := runningWorkers + delta - - // Cannot go below minWorkers - if targetSize < p.minWorkers { - targetSize = p.minWorkers - } - // Cannot go above maxWorkers - if targetSize > p.maxWorkers { - targetSize = p.maxWorkers - } +// startWorkers creates new worker goroutines to run the given tasks +func (p *WorkerPool) startWorker(firstTask func()) { - if p.debug { - // Print debugging information - durationSecs := duration.Seconds() - inputRate := float64(incomingTasks) / durationSecs - outputRate := float64(completedTasks) / durationSecs - message := fmt.Sprintf("%d\t%d\t%d\t%d\t\"%f\"\t\"%f\"\t%d\t\"%f\"\n", - runningWorkers, idleWorkers, incomingTasks, completedTasks, - inputRate, outputRate, - delta, durationSecs) - fmt.Printf(message) - } + // Increment worker count + p.incrementWorkerCount() - return targetSize - runningWorkers + // Launch worker + go worker(firstTask, p.tasks, &p.idleWorkerCount, p.decrementWorkerCount, p.panicHandler) } -// startWorkers creates new worker goroutines to run the given tasks -func (p *WorkerPool) startWorkers(count int, firstTasks []func()) { +func (p *WorkerPool) incrementWorkerCount() { // Increment worker count - workerCount := atomic.AddInt32(&p.workerCount, int32(count)) - - // Collect debug information - if p.debug && int(workerCount) > p.maxWorkerCount { - p.maxWorkerCount = int(workerCount) - } + atomic.AddInt32(&p.workerCount, 1) // Increment waiting group semaphore - p.waitGroup.Add(count) - - // Launch workers - var firstTask func() - for i := 0; i < count; i++ { - firstTask = nil - if i < len(firstTasks) { - firstTask = firstTasks[i] - } - go worker(firstTask, p.dispatchedTasks, &p.idleWorkerCount, &p.completedTaskCount, p.decrementWorkers, p.panicHandler) - } + p.waitGroup.Add(1) } -func (p *WorkerPool) decrementWorkers() { +func (p *WorkerPool) decrementWorkerCount() { // Decrement worker count atomic.AddInt32(&p.workerCount, -1) @@ -464,32 +300,31 @@ func (p *WorkerPool) Group() *TaskGroup { } // worker launches a worker goroutine -func worker(firstTask func(), tasks chan func(), idleWorkerCount *int32, completedTaskCount *uint64, exitHandler func(), panicHandler func(interface{})) { +func worker(firstTask func(), tasks chan func(), idleWorkerCount *int32, exitHandler func(), panicHandler func(interface{})) { defer func() { + if panic := recover(); panic != nil { // Handle panic panicHandler(panic) // Restart goroutine - go worker(nil, tasks, idleWorkerCount, completedTaskCount, exitHandler, panicHandler) + go worker(nil, tasks, idleWorkerCount, exitHandler, panicHandler) } else { - // Handle exit + // Handle normal exit exitHandler() + + // Decrement idle count + atomic.AddInt32(idleWorkerCount, -1) } }() // We have received a task, execute it - func() { - // Increment idle count - defer atomic.AddInt32(idleWorkerCount, 1) - if firstTask != nil { - // Increment completed task count - defer atomic.AddUint64(completedTaskCount, 1) - - firstTask() - } - }() + if firstTask != nil { + firstTask() + } + // Increment idle count + atomic.AddInt32(idleWorkerCount, 1) for task := range tasks { if task == nil { @@ -501,15 +336,10 @@ func worker(firstTask func(), tasks chan func(), idleWorkerCount *int32, complet atomic.AddInt32(idleWorkerCount, -1) // We have received a task, execute it - func() { - // Increment idle count - defer atomic.AddInt32(idleWorkerCount, 1) - - // Increment completed task count - defer atomic.AddUint64(completedTaskCount, 1) + task() - task() - }() + // Increment idle count + atomic.AddInt32(idleWorkerCount, 1) } } diff --git a/pond_blackbox_test.go b/pond_blackbox_test.go index 3485e8e..f9599bb 100644 --- a/pond_blackbox_test.go +++ b/pond_blackbox_test.go @@ -164,6 +164,32 @@ func TestSubmitBeforeWithNilTask(t *testing.T) { assertEqual(t, 0, pool.Running()) } +func TestTrySubmit(t *testing.T) { + + pool := pond.New(1, 5) + + // Submit a long-running task + var doneCount int32 + pool.Submit(func() { + time.Sleep(5 * time.Millisecond) + atomic.AddInt32(&doneCount, 1) + }) + + // Attempt to submit a task without blocking + dispatched := pool.TrySubmit(func() { + time.Sleep(5 * time.Millisecond) + atomic.AddInt32(&doneCount, 1) + }) + + // Task was not dispatched because the pool was full + assertEqual(t, false, dispatched) + + pool.StopAndWait() + + // Only the first task must have executed + assertEqual(t, int32(1), atomic.LoadInt32(&doneCount)) +} + func TestRunning(t *testing.T) { workerCount := 5 @@ -338,3 +364,26 @@ func TestGroupSubmit(t *testing.T) { assertEqual(t, int32(taskCount), atomic.LoadInt32(&doneCount)) } + +func TestPoolWithCustomStrategy(t *testing.T) { + + pool := pond.New(3, 3, pond.Strategy(pond.RatedResizer(2))) + + // Submit 3 tasks + group := pool.Group() + for i := 0; i < 3; i++ { + group.Submit(func() { + time.Sleep(10 * time.Millisecond) + }) + } + + // Wait for them to complete + group.Wait() + + // 2 workers should have been started + assertEqual(t, 2, pool.Running()) + + pool.StopAndWait() + + assertEqual(t, 0, pool.Running()) +} diff --git a/resizer.go b/resizer.go index ec9666e..d442fad 100644 --- a/resizer.go +++ b/resizer.go @@ -1,190 +1,56 @@ package pond import ( - "container/ring" - "math" - "time" + "runtime" + "sync/atomic" ) +var maxProcs = runtime.GOMAXPROCS(0) + // Preset pool resizing strategies var ( // Eager maximizes responsiveness at the expense of higher resource usage, // which can reduce throughput under certain conditions. // This strategy is meant for worker pools that will operate at a small percentage of their capacity - // most of the time and may occasionally receive bursts of tasks. - Eager = func() ResizingStrategy { return DynamicResizer(1, 0.01) } + // most of the time and may occasionally receive bursts of tasks. It's the default strategy. + Eager = func() ResizingStrategy { return RatedResizer(1) } // Balanced tries to find a balance between responsiveness and throughput. - // It's the default strategy and it's suitable for general purpose worker pools or those + // It's suitable for general purpose worker pools or those // that will operate close to 50% of their capacity most of the time. - Balanced = func() ResizingStrategy { return DynamicResizer(3, 0.01) } + Balanced = func() ResizingStrategy { return RatedResizer(maxProcs / 2) } // Lazy maximizes throughput at the expense of responsiveness. // This strategy is meant for worker pools that will operate close to their max. capacity most of the time. - Lazy = func() ResizingStrategy { return DynamicResizer(5, 0.01) } + Lazy = func() ResizingStrategy { return RatedResizer(maxProcs) } ) -// dynamicResizer implements a configurable dynamic resizing strategy -type dynamicResizer struct { - windowSize int - tolerance float64 - incomingTasks *ring.Ring - completedTasks *ring.Ring - duration *ring.Ring - busyWorkers *ring.Ring +// ratedResizer implements a rated resizing strategy +type ratedResizer struct { + rate int + hits int32 } -// DynamicResizer creates a dynamic resizing strategy that gradually increases or decreases -// the size of the pool to match the rate of incoming tasks (input rate) with the rate of -// completed tasks (output rate). -// windowSize: determines how many cycles to consider when calculating input and output rates. -// tolerance: defines a percentage (between 0 and 1) -func DynamicResizer(windowSize int, tolerance float64) ResizingStrategy { - - if windowSize < 1 { - windowSize = 1 - } - if tolerance < 0 { - tolerance = 0 - } +// RatedResizer creates a resizing strategy which can be configured +// to create workers at a specific rate when the pool has no idle workers. +// rate: determines the number of tasks to receive before creating an extra worker. +// A value of 3 can be interpreted as: "Create a new worker every 3 tasks". +func RatedResizer(rate int) ResizingStrategy { - dynamicResizer := &dynamicResizer{ - windowSize: windowSize, - tolerance: tolerance, + if rate < 1 { + rate = 1 } - dynamicResizer.reset() - return dynamicResizer -} - -func (r *dynamicResizer) reset() { - // Create rings - r.incomingTasks = ring.New(r.windowSize) - r.completedTasks = ring.New(r.windowSize) - r.duration = ring.New(r.windowSize) - r.busyWorkers = ring.New(r.windowSize) - // Initialize with 0s - for i := 0; i < r.windowSize; i++ { - r.incomingTasks.Value = 0 - r.completedTasks.Value = 0 - r.duration.Value = 0 * time.Second - r.busyWorkers.Value = 0 - r.incomingTasks = r.incomingTasks.Next() - r.completedTasks = r.completedTasks.Next() - r.duration = r.duration.Next() - r.busyWorkers = r.busyWorkers.Next() + return &ratedResizer{ + rate: rate, } } -func (r *dynamicResizer) totalIncomingTasks() int { - var valueSum int = 0 - r.incomingTasks.Do(func(value interface{}) { - valueSum += value.(int) - }) - return valueSum -} - -func (r *dynamicResizer) totalCompletedTasks() int { - var valueSum int = 0 - r.completedTasks.Do(func(value interface{}) { - valueSum += value.(int) - }) - return valueSum -} - -func (r *dynamicResizer) totalDuration() time.Duration { - var valueSum time.Duration = 0 - r.duration.Do(func(value interface{}) { - valueSum += value.(time.Duration) - }) - return valueSum -} - -func (r *dynamicResizer) avgBusyWorkers() float64 { - var valueSum int = 0 - r.busyWorkers.Do(func(value interface{}) { - valueSum += value.(int) - }) - return float64(valueSum) / float64(r.windowSize) -} - -func (r *dynamicResizer) push(incomingTasks, completedTasks, busyWorkers int, duration time.Duration) { - r.incomingTasks.Value = incomingTasks - r.completedTasks.Value = completedTasks - r.duration.Value = duration - r.busyWorkers.Value = busyWorkers - r.incomingTasks = r.incomingTasks.Next() - r.completedTasks = r.completedTasks.Next() - r.duration = r.duration.Next() - r.busyWorkers = r.busyWorkers.Next() -} - -func (r *dynamicResizer) Resize(runningWorkers, idleWorkers, minWorkers, maxWorkers, incomingTasks, completedTasks int, duration time.Duration) int { - - r.push(incomingTasks, completedTasks, runningWorkers-idleWorkers, duration) - - windowIncomingTasks := r.totalIncomingTasks() - windowCompletedTasks := r.totalCompletedTasks() - windowSecs := r.totalDuration().Seconds() - windowInputRate := float64(windowIncomingTasks) / windowSecs - windowOutputRate := float64(windowCompletedTasks) / windowSecs - - if runningWorkers == 0 || windowCompletedTasks == 0 { - // No workers yet, create as many workers ar.incomingTasks-idleWorkers - delta := incomingTasks - idleWorkers - if delta < 0 { - delta = 0 - } - return r.fitDelta(delta, runningWorkers, minWorkers, maxWorkers) - } +func (r *ratedResizer) Resize(runningWorkers, minWorkers, maxWorkers int) bool { - // Calculate max throughput - avgBusyWorkers := r.avgBusyWorkers() - if avgBusyWorkers < 1 { - avgBusyWorkers = 1 - } - windowWorkerRate := windowOutputRate / avgBusyWorkers - if windowWorkerRate < 1 { - windowWorkerRate = 1 + if r.rate == 1 { + return true } - maxOutputRate := windowWorkerRate * float64(runningWorkers) - deltaRate := windowInputRate - maxOutputRate - - // No changes, do not resize - if deltaRate == 0 { - return 0 - } + hits := int(atomic.AddInt32(&r.hits, 1)) - // If delta % is below the defined tolerance, do not resize - if r.tolerance > 0 { - deltaPercentage := math.Abs(deltaRate / windowInputRate) - if deltaPercentage < r.tolerance { - return 0 - } - } - - if deltaRate > 0 { - ratio := windowSecs / float64(r.windowSize) - delta := int(ratio * (deltaRate / windowWorkerRate)) - if delta < 0 { - delta = 0 - } - if deltaRate > 0 && delta < 1 { - delta = 1 - } - return r.fitDelta(delta, runningWorkers, minWorkers, maxWorkers) - } else if deltaRate < 0 && idleWorkers > 0 { - // Need to shrink the pool - return r.fitDelta(-1, runningWorkers, minWorkers, maxWorkers) - } - return 0 -} - -func (r *dynamicResizer) fitDelta(delta, current, min, max int) int { - if current+delta < min { - delta = -(current - min) - } - if current+delta > max { - delta = max - current - } - return delta + return hits%r.rate == 1 } diff --git a/resizer_test.go b/resizer_test.go index 749feb5..f362468 100644 --- a/resizer_test.go +++ b/resizer_test.go @@ -2,39 +2,43 @@ package pond import ( "testing" - "time" ) -func TestResize(t *testing.T) { +func TestRatedResizer(t *testing.T) { - resizer := DynamicResizer(3, 0.1) + resizer := RatedResizer(3) - // First resize should grow the pool proportionally - assertEqual(t, 10, resizer.Resize(0, 0, 1, 100, 10, 0, 1*time.Second)) + assertEqual(t, true, resizer.Resize(0, 0, 10)) + assertEqual(t, false, resizer.Resize(1, 0, 10)) + assertEqual(t, false, resizer.Resize(2, 0, 10)) + assertEqual(t, true, resizer.Resize(3, 0, 10)) +} - // Now the input rate grows but below the tolerance (10%) - assertEqual(t, -1, resizer.Resize(10, 10, 1, 100, 1, 10, 1*time.Second)) +func TestRatedResizerWithRate1(t *testing.T) { - // Now the input rate grows more - assertEqual(t, 90, resizer.Resize(10, 10, 1, 100, 100000, 11, 1*time.Second)) + resizer := RatedResizer(1) - // Now there's no new tasks for 3 cycles - assertEqual(t, -1, resizer.Resize(10, 10, 1, 100, 0, 100011, 1*time.Second)) - assertEqual(t, -1, resizer.Resize(10, 10, 1, 100, 0, 100011, 1*time.Second)) - assertEqual(t, 0, resizer.Resize(1, 1, 1, 100, 0, 100011, 10*time.Second)) + assertEqual(t, true, resizer.Resize(0, 0, 10)) + assertEqual(t, true, resizer.Resize(1, 0, 10)) + assertEqual(t, true, resizer.Resize(2, 0, 10)) } -func TestEagerPool(t *testing.T) { - pool := New(100, 1000, Strategy(Eager())) - pool.debug = true +func TestRatedResizerWithInvalidRate(t *testing.T) { + + resizer := RatedResizer(0) + + assertEqual(t, true, resizer.Resize(0, 0, 10)) + assertEqual(t, true, resizer.Resize(1, 0, 10)) + assertEqual(t, true, resizer.Resize(2, 0, 10)) +} - for i := 0; i < 100; i++ { - pool.Submit(func() { - time.Sleep(1 * time.Millisecond) - }) - } +func TestPresetRatedResizers(t *testing.T) { - pool.StopAndWait() + eager := Eager() + balanced := Balanced() + lazy := Lazy() - assertEqual(t, 100, pool.maxWorkerCount) + assertEqual(t, true, eager.Resize(0, 0, 10)) + assertEqual(t, true, balanced.Resize(0, 0, 10)) + assertEqual(t, true, lazy.Resize(0, 0, 10)) }