Uneven tabling performance in BProlog 8.1

问题

I did a few experiments with the tabling capabilities of b-prolog version 8.1 and was quite surprised by the performance I observed.

Here is the code that I used. It counts the number of Collatz steps N required for reducing some positive integer I down to 1:

%:- table posInt_CollatzSteps/2.               % remove comment to enable tabling
posInt_CollatzSteps(I,N) :-
   (  I == 1
   -> N = 0                                                % base case
   ;  1 is I /\ 1 
   -> I0 is I*3+1, posInt_CollatzSteps(I0,N0), N is N0+1   % odd
   ;  I0 is I>>1,  posInt_CollatzSteps(I0,N0), N is N0+1   % even
   ).

To determine the maximum number of reduction steps required for all integers from I0 to I:

i0_i_maxSteps0_maxSteps(I0,I,M0,M) :-
   (  I0 > I
   -> M0 = M
   ;  posInt_CollatzSteps(I0,N0),
      I1 is I0+1,
      M1 is max(M0,N0),
      i0_i_maxSteps0_maxSteps(I1,I,M1,M)
   ).

When I ran some queries ?- time(i0_i_maxSteps0_maxSteps(1,1000000,0,MaxSteps)). without and with tabling, I observed the following runtimes (in seconds):

w/o tabling: 6.784
with tabling: 2.323, 19.78, 3.089, 3.084, 3.081

By adding :- table posInt_CollatzSteps/2. the queries got 2x faster. Still, I'm puzzled:

The 2nd run is more than 5x slower than the 1st. Apparently most time is spend in GC. From the 3rd run onwards, the tabling variant is fast again.
Warm runs (3rd, 4th,...) are noticeably slower than the cold (1st) run.

I wasn't expecting this! Contrast this with the runtime that I observed with xsb version 3.6.0:

w/o tabling: 14.287
with tabling: 1.829, 0.31, 0.308, 0.31, 0.333

What can I do? Are there any directives or flags to help me get better performance with BProlog? I use BProlog version 8.1 64-bit edition with Linux.

回答1:

Was checking against Jekejeke Prolog trailed memoing. Was only going to n=100'000. For B-Prolog the following command line worked fine on windows:

bp -s 40000000

This is said to amount to a stack/heap space of 160MB. I get both tabled and non-tabled better warm runs than cold runs:

B-Prolog n=100'000 in s:
w/o tabling: 14.276, 12.229
with tabling: 0.419, 0.143, 0.127, 0.152

The memoing code for Jekejeke uses the predicate assumez/2 from the module hypo. In contrast to B-Prolog tabling, the trailed memoing will recalculate everything upon each call. You need to manually add it to your code:

Jekejeke Prolog/Minlog n=100'000 in s:
w/o memoing: 4.521, 3.893
with memoing: 0.724, 0.573, 0.585, 0.555

So there is still room for speed improvements in Jekejeke. Concerning your B-Prolog question: When the memory is too tight, this might irregularily put additional pressure on a GC, maybe resulting in your observed behaviour.

Bye

P.S.: This is the Jekejeke Prolog memoing code. You need to install Jekejeke Minlog to have the module hypo available. Best is to install the latest release:

:- thread_local posInt_CollatzStepsm/2.

mposInt_CollatzSteps(I,N) :-
   (  I == 1
   -> N = 0                                                % base case
   ;  posInt_CollatzStepsm(I,N) %%% memo check
   -> true
   ;  1 is I /\ 1
   -> I0 is I*3+1, mposInt_CollatzSteps(I0,N0), N is N0+1,   % odd
      assumez(posInt_CollatzStepsm(I,N)) %%% memo add
   ;  I0 is I>>1,  mposInt_CollatzSteps(I0,N0), N is N0+1,   % even
      assumez(posInt_CollatzStepsm(I,N)) %%% memo add
   ).

?- time(mi0_i_maxSteps0_maxSteps(1, 100000, 0, R)).
% Up 724 ms, GC 71 ms, Thread Cpu 640 ms (Current 06/20/19 09:43:24)
R = 350

?- time(mi0_i_maxSteps0_maxSteps(1, 100000, 0, R)).
% Up 573 ms, GC 69 ms, Thread Cpu 500 ms (Current 06/20/19 09:43:27)
R = 350

来源：https://stackoverflow.com/questions/30026151/uneven-tabling-performance-in-bprolog-8-1

标签

performance

prolog

memoization

b-prolog

prolog-tabling