After migrating one of my computing heavy backend programs from Java to Go, I find that the performance degraded instead of improving. I tested around some and it seems the array sorting code is the culprit (which I used heavily in my program). I have written the below two simplified programs to do a comparison, and somehow it seems Go's built-in sort function is a lot slower than Java's Arrays.sort method?
package main
import (
"fmt"
"math/rand"
"sort"
"time"
)
func main() {
fmt.Println("Starting")
const x = 1000000
const y = x * 10
var s [y]float64
s1 := rand.NewSource(time.Now().UnixNano())
r1 := rand.New(s1)
start1 := time.Now()
for i := 0; i < y; i++ {
s[i] = r1.Float64()
}
end1 := time.Since(start1)
ss := s[:]
start2 := time.Now()
sort.Float64s(ss)
end2 := time.Since(start2)
fmt.Println(end1)
fmt.Println(end2)
fmt.Println("Number: ", ss[x])
}
and it produces results like this:
Starting
136.6331ms // The time taken to generate 10,000,000 random numbers
3.456781s // The time taken to sort the 10,000,000 random numbers
Number: 0.10000285497001288
While with the Java program here
import java.util.*;
class RSTest {
public static void main(String[] args) {
System.out.println("Starting");
int x = 1000000;
int y = x * 10;
Random gen = new Random(System.currentTimeMillis());
double[] s = new double[y];
long start1 = System.nanoTime();
for (int i = 0; i < y; i++) {
s[i] = gen.nextDouble();
}
long end1 = System.nanoTime();
long start2 = System.nanoTime();
Arrays.sort(s);
long end2 = System.nanoTime();
System.out.println((end1 - start1) / (1000000000.0));
System.out.println((end2 - start2) / (1000000000.0));
System.out.println(s[x]);
}
}
the results are like this
Starting
0.2252634 // The time taken to generate 10,000,000 random numbers
1.0303157 // The time taken to sort the 10,000,000 random numbers
0.0999513608326642
the Go program takes around 130ms to generate 10 million random numbers and assign them to an array while Jave takes around 230ms to generate 10 million random numbers and assign them to an array, this part I think is the improvement I expect from going from Java to Go.
But for the sorting part, it took Java only around 1s to sort the 10 million random numbers but it took Go around 3.5s to do the 10 million random number sort? And this is quite consistent from multiple runs of the test.
So does that mean Go's built-in sort function is really this much inferior to Java's Arrays.sort method? Or did I use Go's sort function wrong? Or something is wrong with my programs?
Thanks.
Note: this is from Go 1.11 and Java 8, the current versions I'm running on my server. Also, please note that the two programs I posted here are purely for testing purposes that I have written in a couple minutes so may (or rather, most certainly do) contain some code that doesn't make much sense for real production systems.
Some Update:
Thanks to @nussjustin's suggestion, I tried sort.Slice with some promising results.
As currently I'm out of office and using a slower notebook, the baseline results for the two above tests are like this now:
For the Java Arrays.sort test
Starting
0.3590694
1.6030528 // The time taken to sort the 10,000,000 random numbers
0.10000905418967532
For the Go sort.Float64s test
Go
Starting
233.1957ms
5.4633992s // The time taken to sort the 10,000,000 random numbers
Number: 0.10002801819954663
And now after modifying the Go test with sort.Slice
package main
import (
"fmt"
"math/rand"
"sort"
"time"
)
func main() {
fmt.Println("Starting")
const x = 1000000
const y = x * 10
var s [y]float64
s1 := rand.NewSource(time.Now().UnixNano())
r1 := rand.New(s1)
start1 := time.Now()
for i := 0; i < y; i++ {
s[i] = r1.Float64()
}
end1 := time.Since(start1)
ss := s[:]
start2 := time.Now()
sort.Slice(ss, func(i, j int) bool { return ss[i] < ss[j] })
end2 := time.Since(start2)
fmt.Println(end1)
fmt.Println(end2)
fmt.Println("Number: ", ss[x])
}
The result is a big improvement over sort.Float64s, but still not as good as Java's array sort
Starting
281.4262ms
3.6745684s // The time taken to sort the 10,000,000 random numbers
Number: 0.10010604106864159
And I think someone complained that there is only 1 distribution for the tests (who later removed his comment), I tested for sorting normal distribution of random numbers too (albeit I'd say such a huge performance difference in sorting uniform distribution of random numbers is already quite a bad sign since the algorithms of sorting uniform distribution of random numbers should be quite mature already)
I just replace the random number generator from uniform distribution to normal distribution like this
Go:
s[i] = r1.NormFloat64()
Java:
s[i] = gen.nextGaussian();
And the result of Java's Arrays.sort method is
Starting
1.4126348
1.6118655
-1.2820310313627319
And Go's sort.Slice
Starting
434.9106ms
3.8936811s
Number: -1.2818667132095363
So Go is sort.Slice is still about twice slower than Java's Arrays.sort, same as for uniform distribution of random numbers. The good thing is that in generating the normal distribution of random numbers, Go is three times faster than Java, compared to about 70% faster in generating uniform distribution of numbers.
Twenty years ago, Java 1.1 was slow. Since then, thousands of people have put their minds to the task to fix that. Today, Java code is usually on par or faster than C++. With Java 12 and GraalVM, we're all going to see another boost.
Bad code in Java can be slow but the same is true for C++. Java doesn't come with a brain, you have to use your own :-)
To answer the question, the code looks correct. My guess is that Java's sort implementation has in fact been optimized to the brink with data from thousands of use cases. Just look at the length: ~3000 lines with tons of corner cases compared to 500 in Go.
thanks for @JimB and @nussjustin's suggestions, I wrote a simple quicksort implementation myself, and it worked the magic!
package main
import (
"fmt"
"math/rand"
"time"
)
func qsort(s []float64) []float64 {
if len(s) < 2 {
return s
}
left, right := 0, len(s)-1
pivot := 0
s[pivot], s[right] = s[right], s[pivot]
for i := range s {
if s[i] < s[right] {
s[left], s[i] = s[i], s[left]
left++
}
}
s[left], s[right] = s[right], s[left]
qsort(s[:left])
qsort(s[left+1:])
return s
}
func main() {
fmt.Println("Starting")
const x = 1000000
const y = x * 10
var s [y]float64
s1 := rand.NewSource(time.Now().UnixNano())
r1 := rand.New(s1)
start1 := time.Now()
for i := 0; i < y; i++ {
s[i] = r1.NormFloat64()
}
end1 := time.Since(start1)
ss := s[:]
start2 := time.Now()
ss = qsort(ss)
end2 := time.Since(start2)
fmt.Println(end1)
fmt.Println(end2)
fmt.Println("Number: ", ss[x])
}
with this super crude quicksort, now I'm able to achieve the following results
Starting
276.763ms
1.589941s
Number: -1.281875446690731
now it's consistently around 15% faster than Java's Arrays.sort method!
I also implemented a quicksort method specifically for array of double in Java to replace Arrays.sort method to see if I can get any performance gain, the performance ends up around the same as Arrays.sort, still around 10% to 15% slower than Go. It seems the Arrays.sort somehow already achieves the best performance possible in Java, and you don't gain anything by stripping away the abstractions.
So I guess the lesson is that if you want performance in Go's sorting, then implement a quicksort function yourself, don't use the built-in sort functions, even the sort.Slice is around twice slower than a self-written sort function, and the sort.Float64s is more than three times (sometimes four times) slower!
I guess these results can finally shut those commenters up about their so-called "invalid benchmark" stuff. Like I said, the performance degradation after migrating from Java to Go is quite real for my production system, and I'd be in a pinch if it cannot be fixed ASAP, now hopefully after replacing all those sort functions, we can finally see a decent performance improvement on our production system so I can have a peace sleep tonight :)