Re: [Bug-apl] Prime Performance and others

[Mike Duvos]

Hi Elias,

It appears I screwed up my previous benchmark of SIEVE.  I apparently forgot I had the XTerm on my desktop logged into an instance of Ubuntu Server LTS on Amazon Web Services, and ended up comparing APL2 on my desktop with GNU APL on a much faster processor.

I've built SVN 669 on my Dell Dimension 2400, an older slower 32 bit machine,

and now get the following results on that box for SIEVE 100000.  4.219 seconds for APL2, and 812.4 seconds for GNU APL.

So GNU APL is about 192.557 times as slow.  There's a bit of variability in the time for GNU APL.  If I put my ear to the box, I can hear the disk being flogged on, and Task Manager shows the memory bouncing around between 8 and 10 meg rapidly, so it's not necessarily true that the processor time is equal to the wall clock time.

I'm not sure how much time you'll get back by reducing redundant array copying.  Any commercial APL will special-case "reshape", "and", and :"index of" on packed boolean vectors, and this is always going to beat GNU APL's object-oriented approach of boxing everything and making no assumptions about the types of values. 

Looking back over the bug list, I see only two unresolved issues.  Building shared libraries under Cygwin/Windows, and specification sometimes giving the wrong answer when you combine mixed structural primitives with bracket selection.

The shared library thing is important, because with AP210 not being completely implemented, Windows users currently have no way to get ASCII files in and out of their workspace.

Regards,

Mike

On Sun, Aug 30, 2015 at 4:20 AM, Elias Mårtenson <address@hidden> wrote:

All right, I have re-run the Sieve benchmark on GNU APL as of svn id 669.

I do see differences, but they are not great. I also ran it on a reduced case in order to have it finish quicker, so the breakdown could be incorrect. I will re-run this on the full (SIEVE 100000) test case and report back.

I can still see Clone taking ridiculous amounts of time though.

Regards,

Elias

On 29 August 2015 at 11:42, Elias Mårtenson <address@hidden> wrote:

I never analysed the Sieve benchmark. I'll take a look at it tonight.

Regards,
Elias

On 29 Aug 2015 06:51, "Mike Duvos" <address@hidden> wrote:

Hi Fred,

I just built SVN 664 on the Dell Dimension I did the original factorial benchmark on.  The time went from 52 seconds to 23.  So that is indeed a dramatic improvement.

I wonder why my Prime Sieve didn't speed up much.

Regards,

Mike

On Fri, Aug 28, 2015 at 11:28 AM, Fred Weigel <address@hidden> wrote:

Juergen

I just built 664 -- the performance of the factorial 300 that Mike Duvos
posted has indeed improved enormously - now running 8 seconds instead of
20 on my reference platform.

In line with expectations from Elias' profiling results.

Very well done!

FredW