Here is the following concurrency example from A Tour of Go
package main
import (
"fmt"
)
func fibonacci(n int, c chan int) {
x, y := 0, 1
for i := 0; i < n; i++ {
c <- x
x, y = y, x+y
}
close(c)
}
func main() {
c := make(chan int, 10)
go fibonacci(cap(c), c)
for i := range c {
fmt.Println(i)
}
}
I modified it to not use goroutines:
package main
import (
"fmt"
)
func fibonacci(n int) int{
if(n==0||n==1){
return 1
}
x:= 1
y:= 1
for i := 0; i < n; i++ {
tmp := x
x = y
y = tmp + y
fmt.Println(x)
}
return x
}
func main(){
fibonacci2(100)
}
However, the time it takes are both nearly instant at n = 100000.
Does anyone have an example where goroutines does speed up calculations? I am wondering if perhaps there are some compiler settings that is limiting the number of cores my program can use.
Why doesn't the goroutines speed up the calculations?
These 2 versions take almost exactly the same time because most of the work is in the Fibonacci function, so it doesn't matter whether it runs on the main goroutine, or a separate goroutine. When n is large, the concurrent version can be slower, because of the communication overhead over channels.
You can see from the above diagram, the only work running on the main thread are 'Println' calls, which take very little time to run.
But if the processing of the number on the main thread takes more time, using a goroutine to generate fibonacci number may be faster.
Related
So I implemented the following prime finding algorithm in go.
primes = []
Assume all numbers are primes (vacuously true)
check = 2
if check is still assumed to be prime append it to primes
multiply check by each prime less than or equal to its minimum factor and
eliminate results from assumed primes.
increment check by 1 and repeat 4 thru 6 until check > limit.
Here is my serial implementation:
package main
import(
"fmt"
"time"
)
type numWithMinFactor struct {
number int
minfactor int
}
func pow(base int, power int) int{
result := 1
for i:=0;i<power;i++{
result*=base
}
return result
}
func process(check numWithMinFactor,primes []int,top int,minFactors []numWithMinFactor){
var n int
for i:=0;primes[i]<=check.minfactor;i++{
n = check.number*primes[i]
if n>top{
break;
}
minFactors[n] = numWithMinFactor{n,primes[i]}
if i+1 == len(primes){
break;
}
}
}
func findPrimes(top int) []int{
primes := []int{}
minFactors := make([]numWithMinFactor,top+2)
check := 2
for power:=1;check <= top;power++{
if minFactors[check].number == 0{
primes = append(primes,check)
minFactors[check] = numWithMinFactor{check,check}
}
process(minFactors[check],primes,top,minFactors)
check++
}
return primes
}
func main(){
fmt.Println("Welcome to prime finder!")
start := time.Now()
fmt.Println(findPrimes(1000000))
elapsed := time.Since(start)
fmt.Println("Finding primes took %s", elapsed)
}
This runs great producing all the primes <1,000,000 in about 63ms (mostly printing) and primes <10,000,000 in 600ms on my pc. Now I figure none of the numbers check such that 2^n < check <= 2^(n+1) have factors > 2^n so I can do all the multiplications and elimination for each check in that range in parallel once I have primes up to 2^n. And my parallel implementation is as follows:
package main
import(
"fmt"
"time"
"sync"
)
type numWithMinFactor struct {
number int
minfactor int
}
func pow(base int, power int) int{
result := 1
for i:=0;i<power;i++{
result*=base
}
return result
}
func process(check numWithMinFactor,primes []int,top int,minFactors []numWithMinFactor, wg *sync.WaitGroup){
defer wg.Done()
var n int
for i:=0;primes[i]<=check.minfactor;i++{
n = check.number*primes[i]
if n>top{
break;
}
minFactors[n] = numWithMinFactor{n,primes[i]}
if i+1 == len(primes){
break;
}
}
}
func findPrimes(top int) []int{
primes := []int{}
minFactors := make([]numWithMinFactor,top+2)
check := 2
var wg sync.WaitGroup
for power:=1;check <= top;power++{
for check <= pow(2,power){
if minFactors[check].number == 0{
primes = append(primes,check)
minFactors[check] = numWithMinFactor{check,check}
}
wg.Add(1)
go process(minFactors[check],primes,top,minFactors,&wg)
check++
if check>top{
break;
}
}
wg.Wait()
}
return primes
}
func main(){
fmt.Println("Welcome to prime finder!")
start := time.Now()
fmt.Println(findPrimes(1000000))
elapsed := time.Since(start)
fmt.Println("Finding primes took %s", elapsed)
}
Unfortunately not only is this implementation slower running up to 1,000,000 in 600ms and up to 10 million in 6 seconds. My intuition tells me that there is potential for parallelism to improve performance however I clearly haven't been able to achieve that and would greatly appreciate any input on how to improve runtime here, or more specifically any insight as to why the parallel solution is slower.
Additionally the parallel solution consumes more memory relative to the serial solution but that is to be expected; the serial solution can grid up to 1,000,000,000 in about 22 seconds where the parallel solution runs out of memory on my system (32GB ram) going for the same target. But I'm asking about runtime here not memory use, I could for example use the zero value state of the minFactors array rather than a separate isPrime []bool true state but I think it is more readable as is.
I've tried passing a pointer for primes []int but that didn't seem to make a difference, using a channel instead of passing the minFactors array to the process function resulted in big time memory use and a much(10x ish) slower performance. I've re-written this algo a couple times to see if I could iron anything out but no luck. Any insights or suggestions would be much appreciated because I think parallelism could make this faster not 10x slower!
Par #Volker's suggestion I limited the number of processes to somthing less than my pc's available logical processes with the following revision however I am still getting runtimes that are 10x slower than the serial implementation.
package main
import(
"fmt"
"time"
"sync"
)
type numWithMinFactor struct {
number int
minfactor int
}
func pow(base int, power int) int{
result := 1
for i:=0;i<power;i++{
result*=base
}
return result
}
func process(check numWithMinFactor,primes []int,top int,minFactors []numWithMinFactor, wg *sync.WaitGroup){
defer wg.Done()
var n int
for i:=0;primes[i]<=check.minfactor;i++{
n = check.number*primes[i]
if n>top{
break;
}
minFactors[n] = numWithMinFactor{n,primes[i]}
if i+1 == len(primes){
break;
}
}
}
func findPrimes(top int) []int{
primes := []int{}
minFactors := make([]numWithMinFactor,top+2)
check := 2
nlogicalProcessors := 20
var wg sync.WaitGroup
var twoPow int
for power:=1;check <= top;power++{
twoPow = pow(2,power)
for check <= twoPow{
for nLogicalProcessorsInUse := 0 ; nLogicalProcessorsInUse < nlogicalProcessors; nLogicalProcessorsInUse++{
if minFactors[check].number == 0{
primes = append(primes,check)
minFactors[check] = numWithMinFactor{check,check}
}
wg.Add(1)
go process(minFactors[check],primes,top,minFactors,&wg)
check++
if check>top{
break;
}
if check>twoPow{
break;
}
}
wg.Wait()
if check>top{
break;
}
}
}
return primes
}
func main(){
fmt.Println("Welcome to prime finder!")
start := time.Now()
fmt.Println(findPrimes(10000000))
elapsed := time.Since(start)
fmt.Println("Finding primes took %s", elapsed)
}
tldr; Why is my parallel implementation slower than serial implementation how do I make it faster?
Par #mh-cbon's I made larger jobs for parallel processing resulting in the following code.
package main
import(
"fmt"
"time"
"sync"
)
func pow(base int, power int) int{
result := 1
for i:=0;i<power;i++{
result*=base
}
return result
}
func process(check int,primes []int,top int,minFactors []int){
var n int
for i:=0;primes[i]<=minFactors[check];i++{
n = check*primes[i]
if n>top{
break;
}
minFactors[n] = primes[i]
if i+1 == len(primes){
break;
}
}
}
func processRange(start int,end int,primes []int,top int,minFactors []int, wg *sync.WaitGroup){
defer wg.Done()
for start <= end{
process(start,primes,top,minFactors)
start++
}
}
func findPrimes(top int) []int{
primes := []int{}
minFactors := make([]int,top+2)
check := 2
nlogicalProcessors := 10
var wg sync.WaitGroup
var twoPow int
var start int
var end int
var stepSize int
var stepsTaken int
for power:=1;check <= top;power++{
twoPow = pow(2,power)
stepSize = (twoPow-start)/nlogicalProcessors
stepsTaken = 0
stepSize = (twoPow/2)/nlogicalProcessors
for check <= twoPow{
start = check
end = check+stepSize
if stepSize == 0{
end = twoPow
}
if stepsTaken == nlogicalProcessors-1{
end = twoPow
}
if end>top {
end = top
}
for check<=end {
if minFactors[check] == 0{
primes = append(primes,check)
minFactors[check] = check
}
check++
}
wg.Add(1)
go processRange(start,end,primes,top,minFactors,&wg)
if check>top{
break;
}
if check>twoPow{
break;
}
stepsTaken++
}
wg.Wait()
if check>top{
break;
}
}
return primes
}
func main(){
fmt.Println("Welcome to prime finder!")
start := time.Now()
fmt.Println(findPrimes(1000000))
elapsed := time.Since(start)
fmt.Println("Finding primes took %s", elapsed)
}
This runs at a similar speed to the serial implementation.
So I did eventually get a parallel version of the code to run slightly faster than the serial version. following suggestions from #mh-cbon (See above). However this implementation did not result in vast improvements relative to the serial implementation (50ms to 10 million compared to 75ms serially) Considering that allocating and writing an []int 0:10000000 takes 25ms I'm not disappointed by these results. As #Volker stated "such stuff often is not limited by CPU but by memory bandwidth." which I believe is the case here.
I would still love to see any additional improvements however I am somewhat satisfied with what I've gained here.
Serial code running up to 2 billion 19.4 seconds
Parallel code running up to 2 billion 11.1 seconds
Initializing []int{0:2Billion} 4.5 seconds
I'm running a function in a goroutine each time a for-loop iterates, and I'm using sync.WaitGroup to make sure the goroutines all finish. However, I'm getting weird behavior testing the concurrency with counters. In the example below, I attempt to keep track of the thread count using 4 different techniques (w, x, y, z), and get 4 different results. The only result I understand is x, since it is incremented in the for-loop itself. What am I missing here?
package main
import "fmt"
import "sync"
var w = 0
func main() {
x := 0
y := 0
z := 0
var wg sync.WaitGroup
for i := 0; i < 10000; i++ {
wg.Add(1)
x++
go func() {
z++
test(&y)
wg.Done()
}()
}
wg.Wait()
fmt.Println(w, x, y, z) // 8947 10000 8831 8816
}
func test(y *int) {
w++
*y++
}
The sync.Waitgroup is working as expected. w, y and z will not reach 10000 because multiple goroutines are incrementing them concurrently, and Go's increment is not concurrent-safe: it is implemented as a normal fetch-increment-reassign operation.
You have two options.
option 1: mutex
type incrementer struct {
sync.Mutex
i int
}
func (i *incrementer) Add(n int) {
i.Lock()
defer i.Unlock()
i.i += n
}
and use this type for w, y and z.
Full example: https://play.golang.org/p/6wWUK2xnOCW
option 2: sync.atomic
var w int32 = 0
go func(){
// in the loop
atomic.AddInt32(&w, 1)
}()
Full example: https://play.golang.org/p/oUCGgKYC1-Y
I have two versions of factorial. Concurrent vs Sequencial.
Both the program will calculate factorial of 10 "1000000" times.
Factorial Concurrent Processing
package main
import (
"fmt"
//"math/rand"
"sync"
"time"
//"runtime"
)
func main() {
start := time.Now()
printFact(fact(gen(1000000)))
fmt.Println("Current Time:", time.Now(), "Start Time:", start, "Elapsed Time:", time.Since(start))
panic("Error Stack!")
}
func gen(n int) <-chan int {
c := make(chan int)
go func() {
for i := 0; i < n; i++ {
//c <- rand.Intn(10) + 1
c <- 10
}
close(c)
}()
return c
}
func fact(in <-chan int) <-chan int {
out := make(chan int)
var wg sync.WaitGroup
for n := range in {
wg.Add(1)
go func(n int) {
//temp := 1
//for i := n; i > 0; i-- {
// temp *= i
//}
temp := calcFact(n)
out <- temp
wg.Done()
}(n)
}
go func() {
wg.Wait()
close(out)
}()
return out
}
func printFact(in <-chan int) {
//for n := range in {
// fmt.Println("The random Factorial is:", n)
//}
var i int
for range in {
i ++
}
fmt.Println("Count:" , i)
}
func calcFact(c int) int {
if c == 0 {
return 1
} else {
return calcFact(c-1) * c
}
}
//###End of Factorial Concurrent
Factorial Sequencial Processing
package main
import (
"fmt"
//"math/rand"
"time"
"runtime"
)
func main() {
start := time.Now()
//for _, n := range factorial(gen(10000)...) {
// fmt.Println("The random Factorial is:", n)
//}
var i int
for range factorial(gen(1000000)...) {
i++
}
fmt.Println("Count:" , i)
fmt.Println("Current Time:", time.Now(), "Start Time:", start, "Elapsed Time:", time.Since(start))
}
func gen(n int) []int {
var out []int
for i := 0; i < n; i++ {
//out = append(out, rand.Intn(10)+1)
out = append(out, 10)
}
println(len(out))
return out
}
func factorial(val ...int) []int {
var out []int
for _, n := range val {
fa := calcFact(n)
out = append(out, fa)
}
return out
}
func calcFact(c int) int {
if c == 0 {
return 1
} else {
return calcFact(c-1) * c
}
}
//###End of Factorial sequential processing
My assumption was concurrent processing will be faster than sequential but sequential is executing faster than concurrent in my windows machine.
I am using 8 core/ i7 / 32 GB RAM.
I am not sure if there is something wrong in the programs or my basic understanding is correct.
p.s. - I am new to GoLang.
Concurrent version of your program will always be slow compared to the sequential version. The reason however, is related to the nature and behavior of problem you are trying to solve.
Your program is concurrent but it is not parallel. Each callFact is running in it's own goroutine but there is no division of the amount of work required to be done. Each goroutine must perform the same computation and output the same value.
It is like having a task that requires some text to be copied a hundred times. You have just one CPU (ignore the cores for now).
When you start a sequential process, you point the CPU to the original text once, and ask it to write it down a 100 times. The CPU has to manage a single task.
With goroutines, the CPU is told that there are a hundred tasks that must be done concurrently. It just so happens that they are all the same tasks. But CPU is not smart enough to know that.
So it does the same thing as above. Even though each task now is a 100 times smaller, there is still just one CPU. So the amount of work CPU has to do is still the same, except with all the added overhead of managing 100 different things at once. Hence, it looses a part of its efficiency.
To see an improvement in performance you'll need proper parallelism. A simple example would be to split the factorial input number roughly in the middle and compute 2 smaller factorials. Then combine them together:
// not an ideal solution
func main() {
ch := make(chan int)
r := 10
result := 1
go fact(r, ch)
for i := range ch {
result *= i
}
fmt.Println(result)
}
func fact(n int, ch chan int) {
p := n/2
q := p + 1
var wg sync.WaitGroup
wg.Add(2)
go func() {
ch <- factPQ(1, p)
wg.Done()
}()
go func() {
ch <- factPQ(q, n)
wg.Done()
}()
go func() {
wg.Wait()
close(ch)
}()
}
func factPQ(p, q int) int {
r := 1
for i := p; i <= q; i++ {
r *= i
}
return r
}
Working code: https://play.golang.org/p/xLHAaoly8H
Now you have two goroutines working towards the same goal and not just repeating the same calculations.
Note about CPU cores:
In your original code, the sequential version's operations are most definitely being distributed amongst various CPU cores by the runtime environment and the OS. So it still has parallelism to a degree, you just don't controll it.
The same is happening in the concurrent version but again as mentioned above, the overhead of goroutine context switching makes the performance come down.
abhink has given a good answer. I would also like to draw attention to Amdahl's Law, which should always be borne in mind when trying to use parallel processing to increase the overall speed of computation. That's not to say "don't make things parallel", but rather: be realistic about expectations and understand the parallel architecture fully.
Go allows us to write concurrent programs. This is related to trying to write faster parallel programs, but the two issues are separate. See Rob Pike's Concurrency is not Parallelism for more info.
I'll use a hacky inefficient prime number finder to make this question a little more concrete.
Let's say our main function fires off a bunch of "worker" goroutines. They will report their results to a single channnel which prints them. But not every worker will report something so we can't use a counter to know when the last job is finished. Or is there a way?
For the concrete example, here, main fires off goroutines to check whether the values 2...1000 are prime (yeah I know it is inefficient).
package main
import (
"fmt"
"time"
)
func main() {
c := make(chan int)
go func () {
for {
fmt.Print(" ", <- c)
}
}()
for n := 2; n < 1000; n++ {
go printIfPrime(n, c)
}
time.Sleep(2 * time.Second) // <---- THIS FEELS WRONG
}
func printIfPrime(n int, channel chan int) {
for d := 2; d * d <= n; d++ {
if n % d == 0 {
return
}
}
channel <- n
}
My problem is that I don't know how to reliably stop it at the right time. I tried adding a sleep at the end of main and it works (but it might take too long, and this is no way to write concurrent code!). I would like to know if there was a way to send a stop signal through a channel or something so main can stop at the right time.
The trick here is that I don't know how many worker responses there will be.
Is this impossible or is there a cool trick?
(If there's an answer for this prime example, great. I can probably generalize. Or maybe not. Maybe this is app specific?)
Use a WaitGroup.
The following code uses two WaitGroups. The main function uses wgTest to wait for print_if_prime functions to complete. Once they are done, it closes the channel to break the for loop in the printing goroutine. The main function uses wgPrint to wait for printing to complete.
package main
import (
"fmt"
"sync"
)
func main() {
c := make(chan int)
var wgPrint, wgTest sync.WaitGroup
wgPrint.Add(1)
go func(wg *sync.WaitGroup) {
defer wg.Done()
for n := range c {
fmt.Print(" ", n)
}
}(&wgPrint)
for n := 2; n < 1000; n++ {
wgTest.Add(1)
go print_if_prime(&wgTest, n, c)
}
wgTest.Wait()
close(c)
wgPrint.Wait()
}
func print_if_prime(wg *sync.WaitGroup, n int, channel chan int) {
defer wg.Done()
for d := 2; d*d <= n; d++ {
if n%d == 0 {
return
}
}
channel <- n
}
playground example
I'm currently working through the Tour of Go, and I thought that goroutines have been used similarly to Python generators, particularly with Question 66. I thought 66 looked complex, so I rewrote it to this:
package main
import "fmt"
func fibonacci(c chan int) {
x, y := 1, 1
for {
c <- x
x, y = y, x + y
}
}
func main() {
c := make(chan int)
go fibonacci(c)
for i := 0; i < 10; i++ {
fmt.Println(<-c)
}
}
This seems to work. A couple of questions:
If I turn up the buffer size on the channel to say, 10, fibonacci would fill up 10 further spots, as quickly as possible, and main would eat up the spots as quickly as it could go. Is this right? This would be more performant than a buffer size of 1 at the expense of memory, correct?
As the channel doesn't get closed by the fibonacci sender, what happens memory-wise when we go out of scope here? My expectation is that once c and go fibonacci is out of scope, the channel and everything on it gets garbage-collected. My gut tells me this is probably not what happens.
Yes, increasing the buffer size might drastically increase the execution speed of your program, because it will reduce the number of context switches. Goroutines aren't garbage-collected, but channels are. In your example, the fibonacci goroutine will run forever (waiting for another goroutine to read from the channel c), and the channel c will never be destroyed, because the fib-goroutine is still using it.
Here is another, sightly different program, which does not leak memory and is imho more similar to Python's generators:
package main
import "fmt"
func fib(n int) chan int {
c := make(chan int)
go func() {
x, y := 0, 1
for i := 0; i <= n; i++ {
c <- x
x, y = y, x+y
}
close(c)
}()
return c
}
func main() {
for i := range fib(10) {
fmt.Println(i)
}
}
Alternatively, if you do not know how many Fibonacci numbers you want to generate, you have to use another quit channel so that you can send the generator goroutine a signal when it should stop. This is whats explained in golang's tutorial https://tour.golang.org/concurrency/4.
I like #tux21b's answer; having the channel created in the fib() function makes the calling code nice and clean. To elaborate a bit, you only need a separate 'quit' channel if there's no way to tell the function when to stop when you call it. If you only ever care about "numbers up to X", you can do this:
package main
import "fmt"
func fib(n int) chan int {
c := make(chan int)
go func() {
x, y := 0, 1
for x < n {
c <- x
x, y = y, x+y
}
close(c)
}()
return c
}
func main() {
// Print the Fibonacci numbers less than 500
for i := range fib(500) {
fmt.Println(i)
}
}
If you want the ability to do either, this is a little sloppy, but I personally like it better than testing the condition in the caller and then signalling a quit through a separate channel:
func fib(wanted func (int, int) bool) chan int {
c := make(chan int)
go func() {
x, y := 0, 1
for i := 0; wanted(i, x); i++{
c <- x
x, y = y, x+y
}
close(c)
}()
return c
}
func main() {
// Print the first 10 Fibonacci numbers
for n := range fib(func(i, x int) bool { return i < 10 }) {
fmt.Println(n)
}
// Print the Fibonacci numbers less than 500
for n := range fib(func(i, x int) bool { return x < 500 }) {
fmt.Println(n)
}
}
I think it just depends on the particulars of a given situation whether you:
Tell the generator when to stop when you create it by
Passing an explicit number of values to generate
Passing a goal value
Passing a function that determines whether to keep going
Give the generator a 'quit' channel, test the values yourself, and tell it to quit when appropriate.
To wrap up and actually answer your questions:
Increasing the channel size would help performance due to fewer context switches. In this trivial example, neither performance nor memory consumption are going to be an issue, but in other situations, buffering the channel is often a very good idea. The memory used by make (chan int, 100) hardly seems significant in most cases, but it could easily make a big performance difference.
You have an infinite loop in your fibonacci function, so the goroutine running it will run (block on c <- x, in this case) forever. The fact that (once c goes out of scope in the caller) you won't ever again read from the channel you share with it doesn't change that. And as #tux21b pointed out, the channel will never be garbage collected since it's still in use. This has nothing to do with closing the channel (the purpose of which is to let the receiving end of the channel know that no more values will be coming) and everything to do with not returning from your function.
You could use closures to simulate a generator. Here is the example from golang.org.
package main
import "fmt"
// fib returns a function that returns
// successive Fibonacci numbers.
func fib() func() int {
a, b := 0, 1
return func() int {
a, b = b, a+b
return a
}
}
func main() {
f := fib()
// Function calls are evaluated left-to-right.
fmt.Println(f(), f(), f(), f(), f())
}
Using channels to emulate Python generators kind of works, but they introduce concurrency where none is needed, and it adds more complication than's probably needed. Here, just keeping the state explicitly is easier to understand, shorter, and almost certainly more efficient. It makes all your questions about buffer sizes and garbage collection moot.
type fibState struct {
x, y int
}
func (f *fibState) Pop() int {
result := f.x
f.x, f.y = f.y, f.x + f.y
return result
}
func main() {
fs := &fibState{1, 1}
for i := 0; i < 10; i++ {
fmt.Println(fs.Pop())
}
}