So I was bored and decided to test how fast pharo is compared to C++. so I tested addition C++ version: #include <stdio.h> int main() { double x=0; while(x<1000000000) { x = x+1; } return 1; } time ./test1 2.84 real 2.84 user 0.00 sys Pharo version: time ./pharo Ephestos.image eval "|x| x := 0. 1 to: 1000000000 do:[:each| x := x +1]" 1 2.09 real 1.94 user 0.08 sys Pharo is +50% faster than C++ ... o_O ... that's suprising, I assume here Pharo VM probably does some kind of optimisation.
in multiplication pharo is around 2 times slower compared to C++ #include <stdio.h> int main() { double x=1; for(double i; i<1000000000 ; ++i) { x = 0.1*i; } return 1; } time ./testMul 3.13 real 3.13 user 0.00 sys time ./pharo Ephestos.image eval "|x| x := 1. 1 to: 1000000000 do:[:each| x := 0.1 * each]" 1 4.97 real 4.48 user 0.09 sys On Sat, Dec 17, 2016 at 2:16 PM Dimitris Chloupis <kilon.alios@gmail.com> wrote:
So I was bored and decided to test how fast pharo is compared to C++.
so I tested addition
C++ version:
#include <stdio.h>
int main() { double x=0; while(x<1000000000) { x = x+1; }
return 1; }
time ./test1 2.84 real 2.84 user 0.00 sys
Pharo version:
time ./pharo Ephestos.image eval "|x| x := 0. 1 to: 1000000000 do:[:each| x := x +1]" 1 2.09 real 1.94 user 0.08 sys
Pharo is +50% faster than C++ ... o_O ... that's suprising, I assume here Pharo VM probably does some kind of optimisation.
Hmm, this sounds wrong (C++ should be (a lot) faster on such micro benchmarks). You should exclude the executable startup time. In Pharo you should do [ .. ] timeToRun and something similar in C/C++.
On 17 Dec 2016, at 13:41, Dimitris Chloupis <kilon.alios@gmail.com> wrote:
in multiplication pharo is around 2 times slower compared to C++
#include <stdio.h>
int main() { double x=1; for(double i; i<1000000000 ; ++i) { x = 0.1*i; }
return 1; }
time ./testMul 3.13 real 3.13 user 0.00 sys
time ./pharo Ephestos.image eval "|x| x := 1. 1 to: 1000000000 do:[:each| x := 0.1 * each]" 1 4.97 real 4.48 user 0.09 sys
On Sat, Dec 17, 2016 at 2:16 PM Dimitris Chloupis <kilon.alios@gmail.com> wrote: So I was bored and decided to test how fast pharo is compared to C++.
so I tested addition
C++ version:
#include <stdio.h>
int main() { double x=0; while(x<1000000000) { x = x+1; }
return 1; }
time ./test1 2.84 real 2.84 user 0.00 sys
Pharo version:
time ./pharo Ephestos.image eval "|x| x := 0. 1 to: 1000000000 do:[:each| x := x +1]" 1 2.09 real 1.94 user 0.08 sys
Pharo is +50% faster than C++ ... o_O ... that's suprising, I assume here Pharo VM probably does some kind of optimisation.
Like this: [ | sum | sum:= 0. 1 to: 1e9 do: [ :i | sum := sum + i ]. sum ] timeToRun. "0:00:00:39.305" cat bench.c #include <stdio.h> #include <time.h> int main() { clock_t begin = clock(); int sum = 0; for (int i = 0; i < 1000000000; i++) sum += i; clock_t end = clock(); double time_spent = (double)(end - begin) / CLOCKS_PER_SEC; printf("time_spent=%f\n", time_spent); } ./a.out time_spent=2.370234 That is C being 16 times faster ...
On 17 Dec 2016, at 13:49, Sven Van Caekenberghe <sven@stfx.eu> wrote:
Hmm, this sounds wrong (C++ should be (a lot) faster on such micro benchmarks).
You should exclude the executable startup time.
In Pharo you should do [ .. ] timeToRun and something similar in C/C++.
On 17 Dec 2016, at 13:41, Dimitris Chloupis <kilon.alios@gmail.com> wrote:
in multiplication pharo is around 2 times slower compared to C++
#include <stdio.h>
int main() { double x=1; for(double i; i<1000000000 ; ++i) { x = 0.1*i; }
return 1; }
time ./testMul 3.13 real 3.13 user 0.00 sys
time ./pharo Ephestos.image eval "|x| x := 1. 1 to: 1000000000 do:[:each| x := 0.1 * each]" 1 4.97 real 4.48 user 0.09 sys
On Sat, Dec 17, 2016 at 2:16 PM Dimitris Chloupis <kilon.alios@gmail.com> wrote: So I was bored and decided to test how fast pharo is compared to C++.
so I tested addition
C++ version:
#include <stdio.h>
int main() { double x=0; while(x<1000000000) { x = x+1; }
return 1; }
time ./test1 2.84 real 2.84 user 0.00 sys
Pharo version:
time ./pharo Ephestos.image eval "|x| x := 0. 1 to: 1000000000 do:[:each| x := x +1]" 1 2.09 real 1.94 user 0.08 sys
Pharo is +50% faster than C++ ... o_O ... that's suprising, I assume here Pharo VM probably does some kind of optimisation.
This is not a comparable benchmark as the sum in Smalltalk keeps on growing and ends up doing LargeInteger calculations and thus the speed loss, while the c the int just rolls over. Sum in Smalltalk = 500000000500000000 Sum in C = -1243309312 Best regards, Henrik -----Original Message----- From: Pharo-users [mailto:pharo-users-bounces@lists.pharo.org] On Behalf Of Sven Van Caekenberghe Sent: Saturday, December 17, 2016 2:15 PM To: Any question about pharo is welcome <pharo-users@lists.pharo.org> Subject: Re: [Pharo-users] [Stupid Benchmarks] C++ vs Pharo Like this: [ | sum | sum:= 0. 1 to: 1e9 do: [ :i | sum := sum + i ]. sum ] timeToRun. "0:00:00:39.305" cat bench.c #include <stdio.h> #include <time.h> int main() { clock_t begin = clock(); int sum = 0; for (int i = 0; i < 1000000000; i++) sum += i; clock_t end = clock(); double time_spent = (double)(end - begin) / CLOCKS_PER_SEC; printf("time_spent=%f\n", time_spent); } ./a.out time_spent=2.370234 That is C being 16 times faster ...
On 17 Dec 2016, at 13:49, Sven Van Caekenberghe <sven@stfx.eu> wrote:
Hmm, this sounds wrong (C++ should be (a lot) faster on such micro benchmarks).
You should exclude the executable startup time.
In Pharo you should do [ .. ] timeToRun and something similar in C/C++.
On 17 Dec 2016, at 13:41, Dimitris Chloupis <kilon.alios@gmail.com> wrote:
in multiplication pharo is around 2 times slower compared to C++
#include <stdio.h>
int main() { double x=1; for(double i; i<1000000000 ; ++i) { x = 0.1*i; }
return 1; }
time ./testMul 3.13 real 3.13 user 0.00 sys
time ./pharo Ephestos.image eval "|x| x := 1. 1 to: 1000000000 do:[:each| x := 0.1 * each]" 1 4.97 real 4.48 user 0.09 sys
On Sat, Dec 17, 2016 at 2:16 PM Dimitris Chloupis <kilon.alios@gmail.com> wrote: So I was bored and decided to test how fast pharo is compared to C++.
so I tested addition
C++ version:
#include <stdio.h>
int main() { double x=0; while(x<1000000000) { x = x+1; }
return 1; }
time ./test1 2.84 real 2.84 user 0.00 sys
Pharo version:
time ./pharo Ephestos.image eval "|x| x := 0. 1 to: 1000000000 do:[:each| x := x +1]" 1 2.09 real 1.94 user 0.08 sys
Pharo is +50% faster than C++ ... o_O ... that's suprising, I assume here Pharo VM probably does some kind of optimisation.
On 17 Dec 2016, at 14:28, Henrik Nergaard <henrik.nergaard@uia.no> wrote:
This is not a comparable benchmark as the sum in Smalltalk keeps on growing and ends up doing LargeInteger calculations and thus the speed loss, while the c the int just rolls over.
Sum in Smalltalk = 500000000500000000 Sum in C = -1243309312
Yes, I realised afterwards. Shame on me. Now, in Pharo 64 bits: [ | sum | sum:= 0. 1 to: 1e9 do: [ :i | sum := sum + i ]. self assert: sum = (1e9 * (1e9 + 1) / 2) ] timeToRun. "0:00:00:03.379" (1e9 * (1e9 + 1) / 2) < SmallInteger maxVal. "true" So there the speed difference is way smaller.
Best regards, Henrik
-----Original Message----- From: Pharo-users [mailto:pharo-users-bounces@lists.pharo.org] On Behalf Of Sven Van Caekenberghe Sent: Saturday, December 17, 2016 2:15 PM To: Any question about pharo is welcome <pharo-users@lists.pharo.org> Subject: Re: [Pharo-users] [Stupid Benchmarks] C++ vs Pharo
Like this:
[ | sum | sum:= 0. 1 to: 1e9 do: [ :i | sum := sum + i ]. sum ] timeToRun.
"0:00:00:39.305"
cat bench.c
#include <stdio.h> #include <time.h>
int main() { clock_t begin = clock();
int sum = 0; for (int i = 0; i < 1000000000; i++) sum += i;
clock_t end = clock(); double time_spent = (double)(end - begin) / CLOCKS_PER_SEC; printf("time_spent=%f\n", time_spent); }
./a.out
time_spent=2.370234
That is C being 16 times faster ...
On 17 Dec 2016, at 13:49, Sven Van Caekenberghe <sven@stfx.eu> wrote:
Hmm, this sounds wrong (C++ should be (a lot) faster on such micro benchmarks).
You should exclude the executable startup time.
In Pharo you should do [ .. ] timeToRun and something similar in C/C++.
On 17 Dec 2016, at 13:41, Dimitris Chloupis <kilon.alios@gmail.com> wrote:
in multiplication pharo is around 2 times slower compared to C++
#include <stdio.h>
int main() { double x=1; for(double i; i<1000000000 ; ++i) { x = 0.1*i; }
return 1; }
time ./testMul 3.13 real 3.13 user 0.00 sys
time ./pharo Ephestos.image eval "|x| x := 1. 1 to: 1000000000 do:[:each| x := 0.1 * each]" 1 4.97 real 4.48 user 0.09 sys
On Sat, Dec 17, 2016 at 2:16 PM Dimitris Chloupis <kilon.alios@gmail.com> wrote: So I was bored and decided to test how fast pharo is compared to C++.
so I tested addition
C++ version:
#include <stdio.h>
int main() { double x=0; while(x<1000000000) { x = x+1; }
return 1; }
time ./test1 2.84 real 2.84 user 0.00 sys
Pharo version:
time ./pharo Ephestos.image eval "|x| x := 0. 1 to: 1000000000 do:[:each| x := x +1]" 1 2.09 real 1.94 user 0.08 sys
Pharo is +50% faster than C++ ... o_O ... that's suprising, I assume here Pharo VM probably does some kind of optimisation.
On 17 Dec 2016, at 15:23, Sven Van Caekenberghe <sven@stfx.eu> wrote:
On 17 Dec 2016, at 14:28, Henrik Nergaard <henrik.nergaard@uia.no> wrote:
This is not a comparable benchmark as the sum in Smalltalk keeps on growing and ends up doing LargeInteger calculations and thus the speed loss, while the c the int just rolls over.
Sum in Smalltalk = 500000000500000000 Sum in C = -1243309312
Yes, I realised afterwards. Shame on me.
Now, in Pharo 64 bits:
[ | sum | sum:= 0. 1 to: 1e9 do: [ :i | sum := sum + i ]. self assert: sum = (1e9 * (1e9 + 1) / 2) ] timeToRun.
"0:00:00:03.379"
(1e9 * (1e9 + 1) / 2) < SmallInteger maxVal.
"true"
So there the speed difference is way smaller.
And here is a correct C99 version $ cat bench.c #include <stdio.h> #include <time.h> #include <assert.h> int main() { clock_t begin = clock(); long sum = 0; for (int i = 1; i <= 1000000000; i++) sum += i; assert(sum == ((1000000000*1000000001)/2)); clock_t end = clock(); double time_spent = (double)(end - begin) / CLOCKS_PER_SEC; printf("time_spent=%f\n", time_spent); } prometheus:tmp sven$ ./a.out time_spent=2.393986
Best regards, Henrik
-----Original Message----- From: Pharo-users [mailto:pharo-users-bounces@lists.pharo.org] On Behalf Of Sven Van Caekenberghe Sent: Saturday, December 17, 2016 2:15 PM To: Any question about pharo is welcome <pharo-users@lists.pharo.org> Subject: Re: [Pharo-users] [Stupid Benchmarks] C++ vs Pharo
Like this:
[ | sum | sum:= 0. 1 to: 1e9 do: [ :i | sum := sum + i ]. sum ] timeToRun.
"0:00:00:39.305"
cat bench.c
#include <stdio.h> #include <time.h>
int main() { clock_t begin = clock();
int sum = 0; for (int i = 0; i < 1000000000; i++) sum += i;
clock_t end = clock(); double time_spent = (double)(end - begin) / CLOCKS_PER_SEC; printf("time_spent=%f\n", time_spent); }
./a.out
time_spent=2.370234
That is C being 16 times faster ...
On 17 Dec 2016, at 13:49, Sven Van Caekenberghe <sven@stfx.eu> wrote:
Hmm, this sounds wrong (C++ should be (a lot) faster on such micro benchmarks).
You should exclude the executable startup time.
In Pharo you should do [ .. ] timeToRun and something similar in C/C++.
On 17 Dec 2016, at 13:41, Dimitris Chloupis <kilon.alios@gmail.com> wrote:
in multiplication pharo is around 2 times slower compared to C++
#include <stdio.h>
int main() { double x=1; for(double i; i<1000000000 ; ++i) { x = 0.1*i; }
return 1; }
time ./testMul 3.13 real 3.13 user 0.00 sys
time ./pharo Ephestos.image eval "|x| x := 1. 1 to: 1000000000 do:[:each| x := 0.1 * each]" 1 4.97 real 4.48 user 0.09 sys
On Sat, Dec 17, 2016 at 2:16 PM Dimitris Chloupis <kilon.alios@gmail.com> wrote: So I was bored and decided to test how fast pharo is compared to C++.
so I tested addition
C++ version:
#include <stdio.h>
int main() { double x=0; while(x<1000000000) { x = x+1; }
return 1; }
time ./test1 2.84 real 2.84 user 0.00 sys
Pharo version:
time ./pharo Ephestos.image eval "|x| x := 0. 1 to: 1000000000 do:[:each| x := x +1]" 1 2.09 real 1.94 user 0.08 sys
Pharo is +50% faster than C++ ... o_O ... that's suprising, I assume here Pharo VM probably does some kind of optimisation.
On 17 Dec 2016, at 15:52, Sven Van Caekenberghe <sven@stfx.eu> wrote:
On 17 Dec 2016, at 15:23, Sven Van Caekenberghe <sven@stfx.eu> wrote:
On 17 Dec 2016, at 14:28, Henrik Nergaard <henrik.nergaard@uia.no> wrote:
This is not a comparable benchmark as the sum in Smalltalk keeps on growing and ends up doing LargeInteger calculations and thus the speed loss, while the c the int just rolls over.
Sum in Smalltalk = 500000000500000000 Sum in C = -1243309312
Yes, I realised afterwards. Shame on me.
Now, in Pharo 64 bits:
[ | sum | sum:= 0. 1 to: 1e9 do: [ :i | sum := sum + i ]. self assert: sum = (1e9 * (1e9 + 1) / 2) ] timeToRun.
"0:00:00:03.379"
(1e9 * (1e9 + 1) / 2) < SmallInteger maxVal.
"true"
So there the speed difference is way smaller.
And here is a correct C99 version
$ cat bench.c #include <stdio.h> #include <time.h> #include <assert.h>
int main() { clock_t begin = clock();
long sum = 0; for (int i = 1; i <= 1000000000; i++) sum += i;
assert(sum == ((1000000000*1000000001)/2));
clock_t end = clock(); double time_spent = (double)(end - begin) / CLOCKS_PER_SEC; printf("time_spent=%f\n", time_spent); }
prometheus:tmp sven$ ./a.out time_spent=2.393986
But, $ cat bench.c #include <stdio.h> #include <time.h> #include <assert.h> int main() { clock_t begin = clock(); long sum = 0L; for (long i = 1; i <= 1000000000; i++) sum += i; assert(sum == ((1000000000L*1000000001L)/2L)); clock_t end = clock(); double time_spent = (double)(end - begin) / CLOCKS_PER_SEC; printf("time_spent=%f\n", time_spent); return (int)sum; } $ cc -O2 bench.c $ ./a.out time_spent=0.000002 Who knows what optimisation was applied ? Maybe the loop wasn't executed at all ... (but I did add the return statement).
Best regards, Henrik
-----Original Message----- From: Pharo-users [mailto:pharo-users-bounces@lists.pharo.org] On Behalf Of Sven Van Caekenberghe Sent: Saturday, December 17, 2016 2:15 PM To: Any question about pharo is welcome <pharo-users@lists.pharo.org> Subject: Re: [Pharo-users] [Stupid Benchmarks] C++ vs Pharo
Like this:
[ | sum | sum:= 0. 1 to: 1e9 do: [ :i | sum := sum + i ]. sum ] timeToRun.
"0:00:00:39.305"
cat bench.c
#include <stdio.h> #include <time.h>
int main() { clock_t begin = clock();
int sum = 0; for (int i = 0; i < 1000000000; i++) sum += i;
clock_t end = clock(); double time_spent = (double)(end - begin) / CLOCKS_PER_SEC; printf("time_spent=%f\n", time_spent); }
./a.out
time_spent=2.370234
That is C being 16 times faster ...
On 17 Dec 2016, at 13:49, Sven Van Caekenberghe <sven@stfx.eu> wrote:
Hmm, this sounds wrong (C++ should be (a lot) faster on such micro benchmarks).
You should exclude the executable startup time.
In Pharo you should do [ .. ] timeToRun and something similar in C/C++.
On 17 Dec 2016, at 13:41, Dimitris Chloupis <kilon.alios@gmail.com> wrote:
in multiplication pharo is around 2 times slower compared to C++
#include <stdio.h>
int main() { double x=1; for(double i; i<1000000000 ; ++i) { x = 0.1*i; }
return 1; }
time ./testMul 3.13 real 3.13 user 0.00 sys
time ./pharo Ephestos.image eval "|x| x := 1. 1 to: 1000000000 do:[:each| x := 0.1 * each]" 1 4.97 real 4.48 user 0.09 sys
On Sat, Dec 17, 2016 at 2:16 PM Dimitris Chloupis <kilon.alios@gmail.com> wrote: So I was bored and decided to test how fast pharo is compared to C++.
so I tested addition
C++ version:
#include <stdio.h>
int main() { double x=0; while(x<1000000000) { x = x+1; }
return 1; }
time ./test1 2.84 real 2.84 user 0.00 sys
Pharo version:
time ./pharo Ephestos.image eval "|x| x := 0. 1 to: 1000000000 do:[:each| x := x +1]" 1 2.09 real 1.94 user 0.08 sys
Pharo is +50% faster than C++ ... o_O ... that's suprising, I assume here Pharo VM probably does some kind of optimisation.
no its not the loop i tried changing the computation and printing the sum the loops is definetly executed and I am also getting *time_spent=0.000001* ok this is insane but i found what is wrong with your code your loop is for (long i = 1; i <= 1000000000; ++i) but my loop is for (int i = 0; i <= 1000000000; ++i) no its not the long that causes the problem its the fact that you star counting from 1 no idea why but if you start counting from 0 it reports sensible numbers in the case of long sum = 0; for (long i = 0; i <= 1000000000; ++i) { sum = 0+i; } it reports the reasonable *time_spent=0.397446* so maybe its some optimisation it does when the counting starts from 1 not executing the loop in its entirety ? On Sat, Dec 17, 2016 at 6:33 PM Sven Van Caekenberghe <sven@stfx.eu> wrote:
On 17 Dec 2016, at 15:52, Sven Van Caekenberghe <sven@stfx.eu> wrote:
On 17 Dec 2016, at 15:23, Sven Van Caekenberghe <sven@stfx.eu> wrote:
On 17 Dec 2016, at 14:28, Henrik Nergaard <henrik.nergaard@uia.no>
wrote:
This is not a comparable benchmark as the sum in Smalltalk keeps on
growing and ends up doing LargeInteger calculations and thus the speed loss, while the c the int just rolls over.
Sum in Smalltalk = 500000000500000000 Sum in C = -1243309312
Yes, I realised afterwards. Shame on me.
Now, in Pharo 64 bits:
[ | sum | sum:= 0. 1 to: 1e9 do: [ :i | sum := sum + i ]. self assert: sum = (1e9 * (1e9 + 1) / 2) ] timeToRun.
"0:00:00:03.379"
(1e9 * (1e9 + 1) / 2) < SmallInteger maxVal.
"true"
So there the speed difference is way smaller.
And here is a correct C99 version
$ cat bench.c #include <stdio.h> #include <time.h> #include <assert.h>
int main() { clock_t begin = clock();
long sum = 0; for (int i = 1; i <= 1000000000; i++) sum += i;
assert(sum == ((1000000000*1000000001)/2));
clock_t end = clock(); double time_spent = (double)(end - begin) / CLOCKS_PER_SEC; printf("time_spent=%f\n", time_spent); }
prometheus:tmp sven$ ./a.out time_spent=2.393986
But, $ cat bench.c #include <stdio.h> #include <time.h> #include <assert.h> int main() { clock_t begin = clock(); long sum = 0L; for (long i = 1; i <= 1000000000; i++) sum += i; assert(sum == ((1000000000L*1000000001L)/2L)); clock_t end = clock(); double time_spent = (double)(end - begin) / CLOCKS_PER_SEC; printf("time_spent=%f\n", time_spent); return (int)sum; } $ cc -O2 bench.c $ ./a.out time_spent=0.000002 Who knows what optimisation was applied ? Maybe the loop wasn't executed at all ... (but I did add the return statement).
Best regards, Henrik
-----Original Message----- From: Pharo-users [mailto:pharo-users-bounces@lists.pharo.org] On Behalf Of Sven Van Caekenberghe Sent: Saturday, December 17, 2016 2:15 PM To: Any question about pharo is welcome <pharo-users@lists.pharo.org> Subject: Re: [Pharo-users] [Stupid Benchmarks] C++ vs Pharo
Like this:
[ | sum | sum:= 0. 1 to: 1e9 do: [ :i | sum := sum + i ]. sum ] timeToRun.
"0:00:00:39.305"
cat bench.c
#include <stdio.h> #include <time.h>
int main() { clock_t begin = clock();
int sum = 0; for (int i = 0; i < 1000000000; i++) sum += i;
clock_t end = clock(); double time_spent = (double)(end - begin) / CLOCKS_PER_SEC; printf("time_spent=%f\n", time_spent); }
./a.out
time_spent=2.370234
That is C being 16 times faster ...
On 17 Dec 2016, at 13:49, Sven Van Caekenberghe <sven@stfx.eu> wrote:
Hmm, this sounds wrong (C++ should be (a lot) faster on such micro benchmarks).
You should exclude the executable startup time.
In Pharo you should do [ .. ] timeToRun and something similar in C/C++.
On 17 Dec 2016, at 13:41, Dimitris Chloupis <kilon.alios@gmail.com> wrote:
in multiplication pharo is around 2 times slower compared to C++
#include <stdio.h>
int main() { double x=1; for(double i; i<1000000000 ; ++i) { x = 0.1*i; }
return 1; }
time ./testMul 3.13 real 3.13 user 0.00 sys
time ./pharo Ephestos.image eval "|x| x := 1. 1 to: 1000000000 do:[:each| x := 0.1 * each]" 1 4.97 real 4.48 user 0.09 sys
On Sat, Dec 17, 2016 at 2:16 PM Dimitris Chloupis < kilon.alios@gmail.com> wrote: So I was bored and decided to test how fast pharo is compared to C++.
so I tested addition
C++ version:
#include <stdio.h>
int main() { double x=0; while(x<1000000000) { x = x+1; }
return 1; }
time ./test1 2.84 real 2.84 user 0.00 sys
Pharo version:
time ./pharo Ephestos.image eval "|x| x := 0. 1 to: 1000000000 do:[:each| x := x +1]" 1 2.09 real 1.94 user 0.08 sys
Pharo is +50% faster than C++ ... o_O ... that's suprising, I assume here Pharo VM probably does some kind of optimisation.
Yes I stand corrected, with unsigned ints and optimisations turned on the time for C++ for multiplication drops from 3.13 to 0.39 thats an almost 10 times speed up , with doubles and max optimisation on its 0.9 Yes I am using Pharo 32 bit. I think so, unless vmLatest now downloads Pharo 64 bit. So the conclusion is that C++ without the max optimisation on is almost as fast as Pharo. Still 10 times diffirence is not bad at all , on the other hand CPython3 on my iMac needs 118 seconds to finish the multiplication and for some strange reason the python 2.7 included with macos just hangs there forever. So Cpython at for me appears around 30-20 times slower... which is kinda weird too. But then Cpython has no JIT VM.
If I am correct, when you do: double x=0; while(x<1000000000) { x = x+1; } C++ compiles the + to double operation. In Pharo, this code: |x| x := 0. 1 to: 1000000000 do:[:each| x := x +1] is using SmallInteger operations. The code you wrote is fully inlined in Cog's JIT (no message sends). Hence you're comparing double operation versus int operations. That's why Pharo is faster. Pharo should be a little bit slower than C++ if it would use int32 because of the 4 Smallinteger checks it requires for the inlined <= and the 2 inlined +. I am a bit surprised the C++ compiler does not entirely remove the loop as it is side-effect free. Are you compiling in O2 ? For *, firstly you're comparing this time double arithmetic vs double arithmetic, where Pharo is slower due to boxing/unboxing, and secondly Cogit does not inline the *, so there is a message send in the loop in Pharo, leading to worse performance. Are you benchmarking on Pharo 64 bits ? Because this kind of double benchmarks should be way faster on the 64 bits VM than the 32 bits one. Anyway, for benchmarking we normally use the bench suite here: http://squeak.org/codespeed/ and micro-benchmarks are always questionable. On Sat, Dec 17, 2016 at 1:41 PM, Dimitris Chloupis <kilon.alios@gmail.com> wrote:
in multiplication pharo is around 2 times slower compared to C++
#include <stdio.h>
int main() { double x=1; for(double i; i<1000000000 ; ++i) { x = 0.1*i; }
return 1; }
time ./testMul 3.13 real 3.13 user 0.00 sys
time ./pharo Ephestos.image eval "|x| x := 1. 1 to: 1000000000 do:[:each| x := 0.1 * each]" 1 4.97 real 4.48 user 0.09 sys
On Sat, Dec 17, 2016 at 2:16 PM Dimitris Chloupis <kilon.alios@gmail.com> wrote:
So I was bored and decided to test how fast pharo is compared to C++.
so I tested addition
C++ version:
#include <stdio.h>
int main() { double x=0; while(x<1000000000) { x = x+1; }
return 1; }
time ./test1 2.84 real 2.84 user 0.00 sys
Pharo version:
time ./pharo Ephestos.image eval "|x| x := 0. 1 to: 1000000000 do:[:each| x := x +1]" 1 2.09 real 1.94 user 0.08 sys
Pharo is +50% faster than C++ ... o_O ... that's suprising, I assume here Pharo VM probably does some kind of optimisation.
The conditions are unfair: the C++ version uses double, whereas the Pharo version uses Integers. Change the C++ version to use int and then report the results :-). Michal On 17.12.2016 13:16, Dimitris Chloupis wrote:
So I was bored and decided to test how fast pharo is compared to C++.
so I tested addition
C++ version:
#include <stdio.h>
int main() { double x=0; while(x<1000000000) { x = x+1; }
return 1; }
time ./test1 2.84 real 2.84 user 0.00 sys
Pharo version:
time ./pharo Ephestos.image eval "|x| x := 0. 1 to: 1000000000 do:[:each| x := x +1]" 1 2.09 real 1.94 user 0.08 sys
Pharo is +50% faster than C++ ... o_O ... that's suprising, I assume here Pharo VM probably does some kind of optimisation.
You are comparing two different types here, double vs a 32 bit integer, and program starting time should be excluded. Something like: In playground: Smalltalk garbageCollect. [ 0 to: 16r0FFFFFFF do: [ :x | ] ] timeToRun. â try switch the 0 to 0.0 to have the iterations be on a float, leads to much slower timeâ C++ (compile with -std=c++11) #include <iostream> #include <chrono> int main(){ using namespace std::chrono; auto start = high_resolution_clock::now(); for(uint32_t x{0}; x <= 0x0FFFFFFF; ++x){ } auto end = high_resolution_clock::now(); auto ms = duration_cast<milliseconds>(end - start).count(); std::cout << "time used: " << ms << "ms" << std::endl; } Best regards, Henrik From: Pharo-users [mailto:pharo-users-bounces@lists.pharo.org] On Behalf Of Dimitris Chloupis Sent: Saturday, December 17, 2016 1:16 PM To: Any question about pharo is welcome <pharo-users@lists.pharo.org> Subject: [Pharo-users] [Stupid Benchmarks] C++ vs Pharo So I was bored and decided to test how fast pharo is compared to C++. so I tested addition C++ version: #include <stdio.h> int main() { double x=0; while(x<1000000000) { x = x+1; } return 1; } time ./test1 2.84 real 2.84 user 0.00 sys Pharo version: time ./pharo Ephestos.image eval "|x| x := 0. 1 to: 1000000000 do:[:each| x := x +1]" 1 2.09 real 1.94 user 0.08 sys Pharo is +50% faster than C++ ... o_O ... that's suprising, I assume here Pharo VM probably does some kind of optimisation.
participants (5)
-
Clément Bera -
Dimitris Chloupis -
Henrik Nergaard -
Michal Balda -
Sven Van Caekenberghe