[rules-users] Improving Drools Memory Performance
Mark Proctor
mproctor at codehaus.org
Wed Jul 21 22:46:17 EDT 2010
On 22/07/2010 03:28, Jevon Wright wrote:
> Hi Mark and Wolfgang,
>
> Thank you for your replies! Comments below.
>
> A bit of background: I am using Drools to take a given EMF model
> instance, and insert new EObjects into the instance, according to the
> given rules. I try to perform inference top-down, so there is more
> than one iteration of insertion - as objects are inserted, the rules
> need to be re-evaluated. If I understand correctly, this means that I
> can't use a stateless session or the sequential option, because the
> working memory is changing with inserted facts.
>
> The rules don't appear to insert directly, because I insert new
> objects into a queue instead (queue.add(object, drools)) - once rule
> evaluation is complete, I insert the contents of the queue into the
> existing working memory and fire all the rules again. I try to prevent
> the rules modifying the working memory directly. This is also why all
> the rules are of the format (x, ..., y, not z => insert z).
>
> This approach has a number of benefits. It finds inconsistencies in
> the rules and means rules have no order, because inserted facts don't
> effect the working memory immediately. It also allows me to detect
> infinite loops, without restricting the number of times a rule can
> fire. This was described in our 2010 paper [1].
>
> I don't think my implementation of this approach is causing the memory
> problem, but I could be wrong.
>
> detail : DetailWire ( (from == source && to == target) || (from ==
> target && to == source) )
> The above is turned effectively into an MVEL statement, you might
> get better performance with a ConditionalElement 'or' as lont as the
> two are mutually exclusive:
>
> ( DetailWire (from == source, to == target ) or
> DetailWire (from == target, to == source) )
>
>
> I thought this was the case. However in this case, you can't bind the
> variable "detail" (the Drools compiler won't accept the syntax), is
> this correct? I think one solution is to split the rule into two
> separate rules for each "or" part (thus a DSL) - I don't want to have
> to expand these rules by hand.
( $d : DetailWire (from == source, to == target ) or
$d : DetailWire (from == target, to == source) )
is valid
>
> And then i'm not sure what it is you are doing in the second two
> rules, but it looks wrong.
> text : InputTextField ( eContainer == form, eval
> (functions.getAutocompleteInputName(attribute).equals(name)) )
> onInput : EventTrigger ( text.onInput == onInput )
> currentInput : Property ( text.currentInput == currentInput )
>
>
> The point of this rule is to select something like the following (from
> an EMF instance):
>
> <child name="form">
> <child xsi:type="InputTextField" name="...">
> <onInput xsi:type="EventTrigger" ... />
> <currentInput xsi:type="Property" ... />
> <events xsi:type="EventTrigger" ... />
> <properties xsi:type="Property" ... />
> </child>
> </child>
IEventTrigger ( text.onInput == onInput ) - is the text here the bound
variable text, or a field on EventTrigger? the logic isn't very clear.
What you have written in java would look like
eventTrigger.getText().getOnInput().equals( eventTrigger)
Is that what you where expecting? that's why we often use the $ prefix
to differentiate fields from variables.
>
> I can't use use 'eContainer', because 'text' can also contain
> EventTriggers in 'text.events'. These bound variables are then
> supposed to be used later within the rule, either to select other
> variables, or as part of the created element.
>
> I am going to try and remove unused bound variables, though. I think I
> will try and write a script to analyse the exported XML for the rules
> to analyse automatically (I have 264 rules written by hand).
>
> Thanks
> Jevon
>
> [1]: J. Wright and J. Dietrich, "Non-Montonic Model Completion in Web
> Application Engineering," in Proceedings of the 21st Australian
> Software Engineering Conference (ASWEC 2010)
> <http://aswec2010.massey.ac.nz/>, Auckland, New Zealand, 2010.
> http://openiaml.org/#completion
>
> 2010/7/16 Mark Proctor <mproctor at codehaus.org
> <mailto:mproctor at codehaus.org>>
>
> detail : DetailWire ( (from == source&& to == target) || (from == target&& to == source) )
> The above is turned effectively into an MVEL statement, you might get better performance with a ConditionalElement 'or' as lont as the
> two are mutually exclusive:
>
> ( DetailWire (from == source, to == target ) or
> DetailWire (from == target, to == source) )
>
> I saw you did this:
> not ( form : InputForm ( eContainer == container, name ==iterator.name <http://iterator.name> ) )
>
> The 'form' is not accessible outside the 'not', and that rule does not need it.
>
> Is this not a bug. You bind "text". And then i'm not sure what it is you are doing in the second two rules, but it looks wrong.
> text : InputTextField ( eContainer == form, eval (functions.getAutocompleteInputName(attribute).equals(name)) )
> onInput : EventTrigger ( text.onInput == onInput
> currentInput : Property ( text.currentInput == currentInput )
>
> It doesn't look like you are updating the session with facts, i.e. it's a stateless session. See if this helps
>
> KnowledgeBaseConfiguration kconf = KnowledgeBaseFactory.newKnowledgeBaseConfiguration();
> kconf.setOption( SequentialOption.YES );
>
> KnowledgeBase kbase = KnowledgeBaseFactory.newKnowledgeBase( kconf );
> final StatelessKnowledgeSession ksession = kbase.newStatelessKnowledgeSession();
> ksession.execute(....);
>
> In the execute you can provie it with a batch of commands to execute, or just a list of objects, up to you. see stateless session for
> more details.
>
> The SequentialOption may help memory, a small mount, if you aren't doing any working memory modifications (insert/modify/update/retract).
>
> Mark
>
>
> On 16/07/2010 04:16, Jevon Wright wrote:
>> Hi again,
>>
>> By removing all of the simple eval()s from my rules, I have cut
>> heap usage by at least an order of magnitude. However this still
>> isn't enough.
>>
>> Since I am trying to reduce the cross-product size (as in SQL), I
>> recall that most SQL implementations have a "DESCRIBE SELECT"
>> query which provides real-time information about the complexity
>> of a given SQL query - i.e. the size of the tables, indexes used,
>> and so on. Is there any such tool available for Drools? Are there
>> any tools which can provide clues as to which rules are using the
>> most memory?
>>
>> Alternatively, I am wondering what kind of benefit I could expect
>> from using materialized views to create summary tables; that is,
>> deriving and inserting additional facts. This would allow Drools
>> to rewrite queries that currently use eval(), but would increase
>> the size of working memory, so would this actually save heap size?
>>
>> To what extent does Drools rewrite queries? Is there any
>> documentation describing the approaches used?
>>
>> Any other ideas on how to reduce heap memory usage? I'd
>> appreciate any ideas :)
>>
>> Thanks
>> Jevon
>>
>>
>> On Mon, Jul 12, 2010 at 5:56 PM, Jevon Wright <jevon at jevon.org
>> <mailto:jevon at jevon.org>> wrote:
>>
>> Hi Wolfgang and Mark,
>>
>> Thank you for your replies! You were correct: my eval() functions
>> could generally be rewritten into Drools directly.
>>
>> I had one function "connectsDetail" that was constraining
>> unidirectional edges, and could be rewritten from:
>> detail : DetailWire ( )
>> eval ( functions.connectsDetail(detail, source, target) )
>>
>> to:
>> detail : DetailWire ( from == source, to == target )
>>
>> Another function, "connects", was constraining bidirectional
>> edges,
>> and could be rewritten from:
>> sync : SyncWire( )
>> eval ( functions.connects(sync, source, target) )
>>
>> to:
>> sync : SyncWire( (from == source && to == target) || (from
>> == target
>> && to == source) )
>>
>> Finally, the "veto" function could be rewritten from:
>> detail : DetailWire ( )
>> eval ( handler.veto(detail) )
>>
>> to:
>> detail : DetailWire ( overridden == false )
>>
>> I took each of these three changes, and evaluated them
>> separately [1].
>> I found that:
>>
>> 1. Inlining 'connectsDetail' made a huge difference - 10-30%
>> faster
>> execution and 50-60% less allocated heap.
>> 2. Inlining 'connects' made very little difference - 10-30%
>> faster
>> execution, but 0-20% more allocated heap.
>> 3. Inlining 'veto' made no difference - no significant change in
>> execution speed or allocated heap.
>>
>> I think I understand why inlining 'connects' would improve
>> heap usage
>> - because the rules essentially have more conditionals?
>>
>> I also understand why 'veto' made no difference - for most of
>> my test
>> models, "overridden" was never true, so adding this
>> conditional was
>> not making the cross product set any smaller.
>>
>> Finally, I also tested simply joining all of the rules
>> together into
>> one file. This happily made no difference at all (although
>> made it
>> more difficult to edit).
>>
>> So I think I can safely conclude that eval() should be used
>> as little
>> as possible - however, this means that the final rules are
>> made more
>> complicated and less human-readable, so a DSL may be best for my
>> common rule patterns in the future.
>>
>> Thanks again!
>> Jevon
>>
>> [1]:
>> http://www.jevon.org/wiki/Improving_Drools_Memory_Performance
>>
>> On Sat, Jul 10, 2010 at 12:28 AM, Wolfgang Laun
>> <wolfgang.laun at gmail.com <mailto:wolfgang.laun at gmail.com>> wrote:
>> > On 9 July 2010 14:14, Mark Proctor <mproctor at codehaus.org
>> <mailto:mproctor at codehaus.org>> wrote:
>> >> You have many objects there that are not constrained;
>> >
>> > I have an inkling that the functions.*() are hiding just
>> these contraints,
>> > It's certainly the wrong way, starting with oodles of node
>> pairs, just to
>> > pick out connected ones by fishing for the connecting edge.
>> And this
>> > is worsened by trying to find two such pairs which meet at some
>> > DomainSource
>> >
>> > Guesswork, hopefully educated ;-)
>> >
>> > -W
>> >
>> >
>> >> if there are
>> >> multiple versions of those objects you are going to get
>> massive amounts
>> >> of cross products. Think in terms of SQL, each pattern you
>> add is like
>> >> an SQL join.
>> >>
>> >> Mark
>> >> On 09/07/2010 09:20, Jevon Wright wrote:
>> >>> Hi everyone,
>> >>>
>> >>> I am working on what appears to be a fairly complex rule
>> base based on
>> >>> EMF. The rules aren't operating over a huge number of
>> facts (less than
>> >>> 10,000 EObjects) and there aren't too many rules (less
>> than 300), but
>> >>> I am having a problem with running out of Java heap space
>> (set at ~400
>> >>> MB).
>> >>>
>> >>> Through investigation, I came to the conclusion that this
>> is due to
>> >>> the design of the rules, rather than the number of facts.
>> The engine
>> >>> uses less memory inserting many facts that use simple
>> rules, compared
>> >>> with inserting few facts that use many rules.
>> >>>
>> >>> Can anybody suggest some tips for reducing heap memory
>> usage in
>> >>> Drools? I don't have a time constraint, only a
>> heap/memory constraint.
>> >>> A sample rule in my project looks like this:
>> >>>
>> >>> rule "Create QueryParameter for target container of
>> DetailWire"
>> >>> when
>> >>> container : Frame( )
>> >>> schema : DomainSchema ( )
>> >>> domainSource : DomainSource ( )
>> >>> instance : DomainIterator( )
>> >>> selectEdge : SelectEdge ( eval (
>> >>> functions.connectsSelect(selectEdge, instance,
>> domainSource )) )
>> >>> schemaEdge : SchemaEdge ( eval (
>> >>> functions.connectsSchema(schemaEdge, domainSource, schema
>> )) )
>> >>> source : VisibleThing ( eContainer == container )
>> >>> target : Frame ( )
>> >>> instanceSet : SetWire (
>> eval(functions.connectsSet(instanceSet,
>> >>> instance, source )) )
>> >>> detail : DetailWire ( )
>> >>> eval ( functions.connectsDetail(detail, source,
>> target ))
>> >>> pk : DomainAttribute ( eContainer == schema,
>> primaryKey == true )
>> >>> not ( queryPk : QueryParameter ( eContainer ==
>> target, name == pk.name <http://pk.name> ) )
>> >>> eval ( handler.veto( detail ))
>> >>>
>> >>> then
>> >>> QueryParameter qp =
>> handler.generatedQueryParameter(detail, target);
>> >>> handler.setName(qp, pk.getName());
>> >>> queue.add(qp, drools); // wraps insert(...)
>> >>>
>> >>> end
>> >>>
>> >>> I try to order the select statements in an order that
>> will reduce the
>> >>> size of the cross-product (in theory), but I also try and
>> keep the
>> >>> rules fairly human readable. I try to avoid comparison
>> operators like
>> >>> < and>. Analysing a heap dump shows that most of the
>> memory is being
>> >>> used in StatefulSession.nodeMemories> PrimitiveLongMap.
>> >>>
>> >>> I am using a StatefulSession; if I understand correctly,
>> I can't use a
>> >>> StatelessSession with sequential mode since I am
>> inserting facts as
>> >>> part of the rules. If I also understand correctly, I'd
>> like the Rete
>> >>> graph to be tall, rather than wide.
>> >>>
>> >>> Some ideas I have thought of include the following:
>> >>> 1. Creating a separate intermediary meta-model to split
>> up the sizes
>> >>> of the rules. e.g. instead of (if A and B and C then
>> insert D), using
>> >>> (if A and B then insert E; if E and C then insert D).
>> >>> 2. Moving eval() statements directly into the Type(...)
>> selectors.
>> >>> 3. Removing eval() statements. Would this allow for
>> better indexing by
>> >>> the Rete algorithm?
>> >>> 4. Reducing the height, or the width, of the class
>> hierarchy of the
>> >>> facts. e.g. Removing interfaces or abstract classes to
>> reduce the
>> >>> possible matches. Would this make a difference?
>> >>> 5. Conversely, increasing the height, or the width, of
>> the class
>> >>> hierarchy. e.g. Adding interfaces or abstract classes to
>> reduce field
>> >>> accessors.
>> >>> 6. Instead of using EObject.eContainer, creating an explicit
>> >>> containment property in all of my EObjects.
>> >>> 7. Creating a DSL that is human-readable, but allows for the
>> >>> automation of some of these approaches.
>> >>> 8. Moving all rules into one rule file, or splitting up
>> rules into
>> >>> smaller files.
>> >>>
>> >>> Is there kind of profiler for Drools that will let me see
>> the size (or
>> >>> the memory usage) of particular rules, or of the memory
>> used after
>> >>> inference? Ideally I'd use this to profile any changes.
>> >>>
>> >>> Thanks for any thoughts or tips! :-)
>> >>>
>> >>> Jevon
>> >>> _______________________________________________
>> >>> rules-users mailing list
>> >>> rules-users at lists.jboss.org
>> <mailto:rules-users at lists.jboss.org>
>> >>> https://lists.jboss.org/mailman/listinfo/rules-users
>> >>>
>> >>>
>> >>
>> >>
>> >> _______________________________________________
>> >> rules-users mailing list
>> >> rules-users at lists.jboss.org
>> <mailto:rules-users at lists.jboss.org>
>> >> https://lists.jboss.org/mailman/listinfo/rules-users
>> >>
>> >
>> > _______________________________________________
>> > rules-users mailing list
>> > rules-users at lists.jboss.org
>> <mailto:rules-users at lists.jboss.org>
>> > https://lists.jboss.org/mailman/listinfo/rules-users
>> >
>>
>>
>>
>> _______________________________________________
>> rules-users mailing list
>> rules-users at lists.jboss.org <mailto:rules-users at lists.jboss.org>
>> https://lists.jboss.org/mailman/listinfo/rules-users
>
>
> _______________________________________________
> rules-users mailing list
> rules-users at lists.jboss.org <mailto:rules-users at lists.jboss.org>
> https://lists.jboss.org/mailman/listinfo/rules-users
>
>
>
> _______________________________________________
> rules-users mailing list
> rules-users at lists.jboss.org
> https://lists.jboss.org/mailman/listinfo/rules-users
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.jboss.org/pipermail/rules-users/attachments/20100722/10a5e446/attachment.html
More information about the rules-users
mailing list