Back

CR Search Benchmarks:

Equipment:  typical, slightly older Intel machines: 2 GHz processor, 1 GB memory
                        running Maple 9.5 or 10

            The current implementation calculates the combinatorics resulting from a clique replacement, checks it against a catalogue to avoid combinatorially equivalent entries, calculates the symmetry group then counts the number of different paths between selected pairs of estranged vertices.

Benchmarks:
            (5,10)                           11,500 complexes                   3000 secs (50 mins) cpu time
            (6,12)                           14,750 complexes                   3000 secs (50 mins) cpu time
(7,14)  (near Cube)        8,701 complexes                   3806 secs (1 hr, 14 mins)
            (7,14)   (near DC)          1,335 complexes                   1235 secs (20.5 mins)
            (8,16)  (near Cube)        4,137 complexes                   4800 secs (1 hr, 20 mins)
            (8,16)  (near DC)           9,881 complexes                 73820 secs (25 hrs)
            (9,18)  (near Cube)        1,613 complexes                   7679 secs (2 hrs, 8 mins)
            (9,18)  (near DC)           1,570 complexes                 74948 secs (26.4 hrs)

Notes:
            The numbers above are significantly affected by catalogue checks to ensure that combinatorially equivalent complexes are not included twice. (For example, the numbers above do not indicate the five dimensional case being significantly faster. This is the result of the catalogue being nearly complete at 11,500 entries.)

            Counting the number of different paths between certain pairs of estranged vertices is by far the most time consuming operation. A higher dimensional search could selectively perform this calculation.

            The current implementation runs into significant memory problems for 10 dimensions, 20 facets. These arise entirely from the path count calculation and I expect we’ll be able to work around them in order to push the dimensions a bit higher. (12 dimensions, 24 facets seems a plausible goal.)
            Currently, Stage 1 is written in Maple though according to some informal time trials, a switch to Matlab -- for example -- would yield a factor (x4) speed improvement and likely an improvement in memory management.

Back