Start goroutines once and use channels to distribute work

eratosthenes_parallel.go

@@ -5,17 +5,24 @@ import (
 	"os"
 	"runtime"
 	"strconv"
-	"sync"
 	"time"
 )
 
-var markGroup sync.WaitGroup
+type work struct {
+	i int64
+	n0, n1 int64
+}
 
-func markMultiples(sieve []uint64, i int64, n0, n1 int64) {
-	for j := n0; j < n1; j += 2 * i {
+type result bool
+
+func markMultiples(sieve []uint64, workCh chan work, resultCh chan result) {
+	for {
+		w := <-workCh
+		for j := w.n0; j < w.n1; j += 2 * w.i {
 			sieve[j>>7] |= 1 << ((j & 0x7F) >> 1)
 		}
-	markGroup.Done()
+		resultCh <- true
+	}
 }
 
 func main() {
@@ -26,8 +33,13 @@ func main() {
 	nCpus := int64(runtime.NumCPU())
 	t0 := time.Now()
 	sieve := make([]uint64, (n>>7)+1)
-	var found int64
+	workCh := make(chan work)
+	resultCh := make(chan result)
 	var i int64
+	for i = 0; i < nCpus; i++ {
+		go markMultiples(sieve, workCh, resultCh)
+	}
+	var found int64
 	i = 2
 	found++
 	for i = 3; i < n; i += 2 {
@@ -41,16 +53,19 @@ func main() {
 			if n / i < i {
 				continue;
 			}
+			workers := 0
 			for j := i * i; j < n; j = jnext {
 				jnext = ((n-j)/(2*i)/p+1)*(2*i) + j
 				if jnext > n {
 					jnext = n
 				}
-				markGroup.Add(1)
-				go markMultiples(sieve, i, j, jnext)
+				workCh <-work{ i, j, jnext }
+				workers++
 				p--
 			}
-			markGroup.Wait()
+			for k  := 0; k < workers; k++ {
+				_ = <-resultCh
+			}
 		}
 	}
 	d := time.Since(t0)