David-
I reran the final tests but included the vec approach
as well. vec is roughly comparable with perl for
performance while PDL with indadd outperforms
the perl list approach somewhere between $N=100
and $N=1000. PDL can be more than 3-6X faster
for large iteration counts. Considering that PDL:
- has the same memory footprint for processing vec
- has much more powerful and flexible data manipulation
- has simpler syntax for accessing elements (NiceSlice!)
- does all of this is without custom XS, PP, or Inline::C
It is arguably the best alternative for this general
class of problems. The output files from my
timing runs and the test code are attached.
--Chris
On Fri, Apr 20, 2012 at 2:52 PM, Chris Marshall <[email protected]> wrote:
> I reran the timings for both cygwin perl and
> asperl on the same computer (without the
> virus scan running) and got more consistent
> answers (without the low-$N anomalies) than
> from yesterday runs.
>
> The performance cross-over between perl and
> PDL is somewhere between $N=100 and
> $N=1000 with PDL being from 3-6X faster
> than perl for $N>10000. I also note that
> cygwin is 2X slower than the native perl
> so there is good reason to make win32
> PDL fully compatable. :-)
>
> $N Cygwin ASperl
> ---------------------------
> 1 0.02 0.02
> 10 0.05 0.08
> 100 0.4 0.6
> 1000 1.9 3.2
> 10000 3.4 6.6
> 100000 3.6 6.5
> 1000000 3.6 6.1
>
> --Chris
>
> On Fri, Apr 20, 2012 at 8:29 AM, David Mertens <[email protected]>
> wrote:
>> No, I'm just not very experienced with Benchmarks. :-)
>>
>> I'll look into it later today.
>>
>> On Apr 19, 2012 4:29 PM, "Chris Marshall" <[email protected]> wrote:
>>>
>>> I notice you are still running a very small
>>> amount of timings. How are your timings
>>> if you change the -1 to -10 in cmpthese?
>>> Is there a specific reason to leave that off
>>> of the update?
>>>
>>> --Chris
>>>
>>>
>>> On Thu, Apr 19, 2012 at 5:11 PM, David Mertens <[email protected]>
>>> wrote:
>>> > Wow, that makes a sizeable difference, especially for larger values of
>>> > $updates_per_round, where it doesn't need to constantly re-allocate
>>> > temporary memory. I've updated the post to reflect that.
>>> >
>>> > Thanks, Chris!
>>> > David
>>> >
>>> >
>>> > On Thu, Apr 19, 2012 at 3:19 PM, Chris Marshall <[email protected]>
>>> > wrote:
>>> >>
>>> >> I don't have a stackoverflow.com account but if I use
>>> >> your original test code, increase the cpu seconds for
>>> >> the benchmark and replace the dataflow increment
>>> >> code by the indadd() routine, I get speeds from 50X
>>> >> for $N=1, down to 1.9X for $N=1000, and then up to
>>> >> 3-4X as the sizes increase to 10000000 where I
>>> >> stopped testing.
>>> >>
>>> >> The key here is to avoid creating and destroying piddles
>>> >> since the computational work involved here is *very*
>>> >> light. In fact, an Inline::PP routine to handle the core
>>> >> of an algorithm of interest (if more than indadd) + the
>>> >> existing PDL should perform very well for this use
>>> >> case---according to the benchmarks on my system:
>>> >>
>>> >> use PDL;
>>> >> use Benchmark qw/cmpthese/;
>>> >>
>>> >> my $updates_per_round = shift || 1;
>>> >>
>>> >> my $N = 1_000_000;
>>> >> my @perl = (0 .. $N - 1);
>>> >> my $pdl = zeroes $N;
>>> >>
>>> >> cmpthese(-10,{
>>> >> perl => sub{
>>> >> $perl[int(rand($N))]++ for (1..$updates_per_round);
>>> >> },
>>> >> pdl => sub{
>>> >> my $to_update = long(random($updates_per_round) * $N);
>>> >> indadd(1,$to_update,$pdl);
>>> >> ## $pdl->index($to_update)++;
>>> >> }
>>> >> });
>>> >>
>>> >>
>>> >> Cheers,
>>> >> Chris
>>> >>
>>> >> On Wed, Apr 18, 2012 at 10:11 AM, David Mertens
>>> >> <[email protected]> wrote:
>>> >> > Hey folks -
>>> >> >
>>> >> > There's a PDL question on stack overflow
>>> >> > (http://stackoverflow.com/questions/9730678/c-like-arrays-in-perl) to
>>> >> > which
>>> >> > I submitted an answer. I believe my answer is better than the
>>> >> > currently
>>> >> > marked best answer, but it looks like the OP either disagrees or
>>> >> > simply
>>> >> > has
>>> >> > not returned to read my answer. I would appreciate if those of you
>>> >> > who
>>> >> > have
>>> >> > stack overflow accounts could read the responses and up-vote what you
>>> >> > think
>>> >> > is the best answer.
>>> >> >
>>> >> > Thanks!
>>> >> > David
>>> >> >
>>> >> > --
>>> >> > "Debugging is twice as hard as writing the code in the first place.
>>> >> > Therefore, if you write the code as cleverly as possible, you are,
>>> >> > by definition, not smart enough to debug it." -- Brian Kernighan
>>> >> >
>>> >> >
>>> >> > _______________________________________________
>>> >> > Perldl mailing list
>>> >> > [email protected]
>>> >> > http://mailman.jach.hawaii.edu/mailman/listinfo/perldl
>>> >> >
>>> >
>>> >
>>> >
>>> >
>>> > --
>>> > "Debugging is twice as hard as writing the code in the first place.
>>> > Therefore, if you write the code as cleverly as possible, you are,
>>> > by definition, not smart enough to debug it." -- Brian Kernighan
>>> >
#!/usr/bin/perl
#
use PDL;
use Benchmark qw/cmpthese/;
my $N = 1_000_000;
my $updates_per_round = shift || 1;
my $vec = "\0" x ($N * 4);
my @perl = (0 .. $N - 1);
my $pdl = sequence(long,$N);
cmpthese(-20,
{
perl => sub{
$perl[int(rand($N))]++ for (1..$updates_per_round);
},
pdl => sub{
my $to_update = long(random($updates_per_round) * $N);
indadd(1,$to_update,$pdl);
},
set => sub{
for (1..$updates_per_round) {
my $index = int(rand($N));
$pdl->set($index,$pdl->at($index)+1);
}
},
vec => sub {
vec($vec, int(rand($N)), 32)++ for (1..$updates_per_round);;
},
}
);
$N = 1
Rate pdl set vec perl
pdl 11530/s -- -88% -98% -98%
set 98826/s 757% -- -86% -86%
vec 714587/s 6098% 623% -- -1%
perl 718834/s 6135% 627% 1% --
$N = 10
Rate pdl set vec perl
pdl 11284/s -- -4% -95% -95%
set 11704/s 4% -- -94% -94%
vec 209793/s 1759% 1693% -- -1%
perl 211958/s 1778% 1711% 1% --
$N = 100
Rate set pdl perl vec
set 1138/s -- -89% -96% -96%
pdl 10034/s 782% -- -62% -62%
perl 26304/s 2212% 162% -- -1%
vec 26654/s 2242% 166% 1% --
$N = 1000
Rate set perl vec pdl
set 114/s -- -96% -96% -98%
perl 2685/s 2245% -- -3% -49%
vec 2775/s 2324% 3% -- -47%
pdl 5281/s 4513% 97% 90% --
$N = 10000
Rate set perl vec pdl
set 11.4/s -- -96% -96% -99%
perl 269/s 2257% -- -1% -72%
vec 273/s 2291% 1% -- -72%
pdl 976/s 8443% 262% 257% --
$N = 100000
Rate set perl vec pdl
set 1.13/s -- -96% -96% -99%
perl 27.2/s 2305% -- -1% -74%
vec 27.6/s 2336% 1% -- -74%
pdl 107/s 9326% 292% 287% --
$N = 1000000
Rate set perl vec pdl
set 0.114/s -- -96% -96% -99%
perl 2.67/s 2251% -- -2% -74%
vec 2.71/s 2288% 2% -- -74%
pdl 10.4/s 9095% 291% 285% --
$N = 1
Rate pdl set vec perl
pdl 22350/s -- -86% -98% -98%
set 162641/s 628% -- -82% -85%
vec 902084/s 3936% 455% -- -15%
perl 1062223/s 4653% 553% 18% --
$N = 10
Rate set pdl vec perl
set 19940/s -- -4% -92% -93%
pdl 20787/s 4% -- -91% -92%
vec 241356/s 1110% 1061% -- -12%
perl 273586/s 1272% 1216% 13% --
$N = 100
Rate set pdl vec perl
set 1973/s -- -90% -93% -94%
pdl 19679/s 897% -- -33% -36%
vec 29222/s 1381% 48% -- -6%
perl 30954/s 1469% 57% 6% --
$N = 1000
Rate set vec perl pdl
set 195/s -- -93% -94% -98%
vec 2944/s 1408% -- -12% -74%
perl 3330/s 1606% 13% -- -71%
pdl 11337/s 5708% 285% 240% --
$N = 10000
Rate set vec perl pdl
set 20.1/s -- -93% -94% -99%
vec 297/s 1377% -- -18% -88%
perl 363/s 1705% 22% -- -85%
pdl 2425/s 11959% 717% 568% --
$N = 100000
Rate set vec perl pdl
set 2.04/s -- -93% -94% -99%
vec 29.0/s 1325% -- -20% -88%
perl 36.5/s 1693% 26% -- -85%
pdl 242/s 11775% 733% 562% --
$N = 1000000
Rate set vec perl pdl
set 0.195/s -- -93% -94% -99%
vec 2.76/s 1312% -- -21% -88%
perl 3.49/s 1691% 27% -- -85%
pdl 23.4/s 11917% 751% 571% --
_______________________________________________
Perldl mailing list
[email protected]
http://mailman.jach.hawaii.edu/mailman/listinfo/perldl
