If you retract and assert stuff, instead of modifying then you do not
need shadow facts - with a enhancment I have in mind this should allow
us to gain some performance. Its something several users have asked for,
and I think the final implementation idea I have will make sense.
Mark
Geoffrey De Smet wrote:
I am not a sure if shadowfacts aren't actually helping instead: I am
using exists() for a couple of rules and with 3.0.4
With 3.0.4 I actually saw 2 exists rules increase the evaluation time
time by 3, so I suspect the shadowfacts-exists hashing has actually
countered that somewhat.
But it should be interesting to see what happens by trying to tweak it
a bit, for example by turning shadow facts off.
Although I agree with Peter that - like the jvm itself - the
out-of-the-box experience should have a higher priority then the
ability to tweak it with lots of parameters.
For taseree-which-needs-a-new-name (drools-solver?:), I have a similar
idea: the many tweaking options (tabu list size, ...) make it hard to
tweak and scale, so I am thinking of some sort of dynamic defaults in
the long run.
As for specifically for the TravelingTournament problem: I've found
out that there's a much smarter way to do define moves, which would
also allow me to remove some rules and greatly reduce the amount of
moves.
I 'll keep in the current (stupid) way, to demonstrate the difference.
Peter Lin wrote, On 2006-12-15 12:19 AM:
:) 10% feel hardly worth the trouble of changing the config. keep in
mind the lazy slobs like me don't bother changing configs.
if it was like 50%, then it would make sense to turn shadow facts
off. Plus, it's already faster than JESS, so not like an user really
needs to turn shadow off.
my bias half bits
peter
On 5/17/06, *Mark Proctor* <[EMAIL PROTECTED]
<mailto:[EMAIL PROTECTED]>> wrote:
Peter,
When we make shadow facts optional we should be able to get
another 10%
speed improvements on there, atleast. You can also play around
with the
vairous indexes, you may find turning off alpha indexing or left and
right indexing speeds things up. We will also make it possible
to turn
off truth maintenance, which should gain a few more percent. I
suspect
there may be more performance improvements, maybe a customised
result
set for faster access to the results (but less easy to use) - you
can
try this yourself by looking at hte QueryTerminalNode and profiling
that. Btw we have free jprofiler licenses for ALL JBoss Rules
contributors, will email you one now.
Mark
Geoffrey De Smet wrote:
> Looks like the trunk is at least 50% faster in my traveling
tournament
> lvl10 use case :)
>
> Here are some rough numbers
> ===========================
>
> Iteration
> Timemillis 3.0.5
> Difference 3.0.5
> Timemillis trunk
> Difference trunk
> Hotspot???
> Start 255671 1107781
> 0 308453 52782 1113359 5578
> 1 320968 12515 1118500 5141
> 2 333531 12563 1123640 5140
> 3 346093 12562 1128765 5125
> 4 358718 12625 1133859 5094
> 5 411156 52438 1138984 5125 3.0.5
> 6 423640 12484 1144046 5062
> 7 436250 12610 1149093 5047
> 8 448703 12453 1154156 5063
> 9 501125 52422 1159203 5047 3.0.5
> 10 513531 12406 1204234 45031 Trunk
> 11 526031 12500 1209265 5031
> 12 538500 12469 1214234 4969
> 13 550843 12343 1219218 4984
> 14 603296 52453 1224250 5032 3.0.5
> 15 615718 12422 1229140 4890
>
> It 15 – 5 204562 90156
>
> I 'll have to run it longer to clear away the hotspot effects
though,
> to get some real good numbers
>
>
---------------------------------------------------------------------
To unsubscribe from this list please visit:
http://xircles.codehaus.org/manage_email
---------------------------------------------------------------------
To unsubscribe from this list please visit:
http://xircles.codehaus.org/manage_email
