Performance for LUBM benchmark

[josd]

The results that we get for the LUBM benchmark are correct but it runs for more than an hour 🤔

The files are at lubm and we also use roundel.pl which is one level up i.e. we load it in rules.pl as :- use_module('../roundel.pl'). and it is to do forward chaining as we otherwise get nontermination and also because there are some existential rules like
'http://www.w3.org/1999/02/22-rdf-syntax-ns#type'(A, 'http://www.example.org/Chair') => 'http://www.example.org/headOf'(A, B), 'http://www.w3.org/1999/02/22-rdf-syntax-ns#type'(B, 'http://www.example.org/Department').

Here is the timing

$ time scryer-prolog -g run,halt rules.pl facts.pl > result.pl

real    69m14.502s
user    69m10.339s
sys     0m4.090s

and here is the timing for Trealla

$ time tpl -g run,halt rules.pl facts.pl > result.pl

real    0m17.687s
user    0m17.667s
sys     0m0.020s

which is about a factor 235 times faster!

What could be the reason for this big gap?
Is there anything that we could improve in our rules?

[mthom]

expand_goal in error.pl strikes again (not only there but many other meta-predicates too)! I'm working on this now. It will be a while though, it's a fairly substantial change. Fortunately I have a week of vacation approaching.

[mthom]

Once again, I was wrong. The call optimizations I've implemented were not that impactful in this benchmark. Rather, retract/1 is dreadfully slow. By replacing retract with bb_get / bb_put and tightening the implementation of assertz, I get a run time of 4m20s on the release compile.

I will work on improving retract/1. The full weight of the WAM compiler is used in assertz so I'm not sure how much can be done to improve on the 4m20s runtime in the immediate future.

[josd]

For the moment I can live with bb_get/2 and bb_put/2 in https://github.com/josd/plows/blob/master/plows.pl
That way https://github.com/josd/plows/blob/master/lubm/test takes indeed sec=233.850 kB=1180236 cpu=99%