perf: tail latency with goSched (#2033)

Alternate implementation to #2016 that doesn't reduce RPS with lower
amounts of threads

caddy/mercure-skip.go

@@ -1,4 +1,5 @@
 //go:build nomercure
+
 package caddy
 
 import (

caddy/workerconfig.go

@@ -39,7 +39,7 @@ type workerConfig struct {
 	// MaxConsecutiveFailures sets the maximum number of consecutive failures before panicking (defaults to 6, set to -1 to never panick)
 	MaxConsecutiveFailures int `json:"max_consecutive_failures,omitempty"`
 
-	requestOptions     []frankenphp.RequestOption
+	requestOptions []frankenphp.RequestOption
 }
 
 func parseWorkerConfig(d *caddyfile.Dispenser) (workerConfig, error) {

frankenphp.go

@@ -307,7 +307,7 @@ func Init(options ...Option) error {
 		return err
 	}
 
-	regularRequestChan = make(chan contextHolder, opt.numThreads-workerThreadCount)
+	regularRequestChan = make(chan contextHolder)
 	regularThreads = make([]*phpThread, 0, opt.numThreads-workerThreadCount)
 	for i := 0; i < opt.numThreads-workerThreadCount; i++ {
 		convertToRegularThread(getInactivePHPThread())

threadregular.go

@@ -2,7 +2,9 @@ package frankenphp
 
 import (
 	"context"
+	"runtime"
 	"sync"
+	"sync/atomic"
 )
 
 // representation of a non-worker PHP thread
@@ -16,9 +18,10 @@ type regularThread struct {
 }
 
 var (
-	regularThreads     []*phpThread
-	regularThreadMu    = &sync.RWMutex{}
-	regularRequestChan chan contextHolder
+	regularThreads       []*phpThread
+	regularThreadMu      = &sync.RWMutex{}
+	regularRequestChan   chan contextHolder
+	queuedRegularThreads = atomic.Int32{}
 )
 
 func convertToRegularThread(thread *phpThread) {
@@ -81,6 +84,7 @@ func (handler *regularThread) waitForRequest() string {
 		// go back to beforeScriptExecution
 		return handler.beforeScriptExecution()
 	case ch = <-regularRequestChan:
+	case ch = <-handler.thread.requestChan:
 	}
 
 	handler.ctx = ch.ctx
@@ -100,23 +104,35 @@ func (handler *regularThread) afterRequest() {
 func handleRequestWithRegularPHPThreads(ch contextHolder) error {
 	metrics.StartRequest()
 
-	select {
-	case regularRequestChan <- ch:
-		// a thread was available to handle the request immediately
-		<-ch.frankenPHPContext.done
-		metrics.StopRequest()
+	runtime.Gosched()
 
-		return nil
-	default:
-		// no thread was available
+	if queuedRegularThreads.Load() == 0 {
+		regularThreadMu.RLock()
+		for _, thread := range regularThreads {
+			select {
+			case thread.requestChan <- ch:
+				regularThreadMu.RUnlock()
+				<-ch.frankenPHPContext.done
+				metrics.StopRequest()
+
+				return nil
+			default:
+				// thread was not available
+			}
+		}
+		regularThreadMu.RUnlock()
 	}
 
 	// if no thread was available, mark the request as queued and fan it out to all threads
+	queuedRegularThreads.Add(1)
 	metrics.QueuedRequest()
+
 	for {
 		select {
 		case regularRequestChan <- ch:
+			queuedRegularThreads.Add(-1)
 			metrics.DequeuedRequest()
+
 			<-ch.frankenPHPContext.done
 			metrics.StopRequest()
 
@@ -125,7 +141,9 @@ func handleRequestWithRegularPHPThreads(ch contextHolder) error {
 			// the request has triggered scaling, continue to wait for a thread
 		case <-timeoutChan(maxWaitTime):
 			// the request has timed out stalling
+			queuedRegularThreads.Add(-1)
 			metrics.DequeuedRequest()
+			metrics.StopRequest()
 
 			ch.frankenPHPContext.reject(ErrMaxWaitTimeExceeded)
 

worker.go

@@ -6,8 +6,10 @@ import (
 	"fmt"
 	"os"
 	"path/filepath"
+	"runtime"
 	"strings"
 	"sync"
+	"sync/atomic"
 	"time"
 
 	"github.com/dunglas/frankenphp/internal/fastabs"
@@ -28,6 +30,7 @@ type worker struct {
 	maxConsecutiveFailures int
 	onThreadReady          func(int)
 	onThreadShutdown       func(int)
+	queuedRequests         atomic.Int32
 }
 
 var (
@@ -253,24 +256,30 @@ func (worker *worker) isAtThreadLimit() bool {
 func (worker *worker) handleRequest(ch contextHolder) error {
 	metrics.StartWorkerRequest(worker.name)
 
-	// dispatch requests to all worker threads in order
-	worker.threadMutex.RLock()
-	for _, thread := range worker.threads {
-		select {
-		case thread.requestChan <- ch:
-			worker.threadMutex.RUnlock()
-			<-ch.frankenPHPContext.done
-			metrics.StopWorkerRequest(worker.name, time.Since(ch.frankenPHPContext.startedAt))
+	runtime.Gosched()
 
-			return nil
-		default:
-			// thread is busy, continue
+	if worker.queuedRequests.Load() == 0 {
+		// dispatch requests to all worker threads in order
+		worker.threadMutex.RLock()
+		for _, thread := range worker.threads {
+			select {
+			case thread.requestChan <- ch:
+				worker.threadMutex.RUnlock()
+				<-ch.frankenPHPContext.done
+				metrics.StopWorkerRequest(worker.name, time.Since(ch.frankenPHPContext.startedAt))
+
+				return nil
+			default:
+				// thread is busy, continue
+			}
 		}
+		worker.threadMutex.RUnlock()
 	}
-	worker.threadMutex.RUnlock()
 
 	// if no thread was available, mark the request as queued and apply the scaling strategy
+	worker.queuedRequests.Add(1)
 	metrics.QueuedWorkerRequest(worker.name)
+
 	for {
 		workerScaleChan := scaleChan
 		if worker.isAtThreadLimit() {
@@ -279,6 +288,7 @@ func (worker *worker) handleRequest(ch contextHolder) error {
 
 		select {
 		case worker.requestChan <- ch:
+			worker.queuedRequests.Add(-1)
 			metrics.DequeuedWorkerRequest(worker.name)
 			<-ch.frankenPHPContext.done
 			metrics.StopWorkerRequest(worker.name, time.Since(ch.frankenPHPContext.startedAt))
@@ -288,7 +298,9 @@ func (worker *worker) handleRequest(ch contextHolder) error {
 			// the request has triggered scaling, continue to wait for a thread
 		case <-timeoutChan(maxWaitTime):
 			// the request has timed out stalling
+			worker.queuedRequests.Add(-1)
 			metrics.DequeuedWorkerRequest(worker.name)
+			metrics.StopWorkerRequest(worker.name, time.Since(ch.frankenPHPContext.startedAt))
 
 			ch.frankenPHPContext.reject(ErrMaxWaitTimeExceeded)