[rules-users] help with diagnosing CEP performance issue
Wolfgang Laun
wolfgang.laun at gmail.com
Sun Dec 26 02:14:54 EST 2010
You are inserting a day's worth of backup events at a time. In this loop the
engine is forced to evaluate both rules, for an ever-increasing number of
events. It's the second rule that requires a comparison between any two
backup events, and this causes a quadratic increment in processing time.
CEP is short for "continuous event processing", and that's what you should
let the engine do, i.e., not only evaluate the conditions but also the
execute the consequences. This is done by calling fireAllRules() after each
insertion. Alternatively, consider running fireUntilHalt() in once thread
and insert in another thread.
Ideally, the test should generate events ordered by time, and the
pseudo-clock might be advanced due to the time stamp in the generated
events. This would give you a very close approximation to a production run.
-W
2010/12/26 radai <radai.rosenblatt at gmail.com>
> well, fixed now :-)
> the benchmark (part of a junit) looks like this:
>
> protected void init() {
> KnowledgeBuilderConfiguration bConf =
> KnowledgeBuilderFactory.newKnowledgeBuilderConfiguration();
> KnowledgeBuilder builder =
> KnowledgeBuilderFactory.newKnowledgeBuilder(bConf);
> builder.add(...);
>
> KnowledgeBaseConfiguration kbConf =
> KnowledgeBaseFactory.newKnowledgeBaseConfiguration();
> kbConf.setOption( EventProcessingOption.STREAM );
> //todo - turn this on when
> https://issues.jboss.org/browse/JBRULES-2845 is fixed
> //kbConf.setOption( MultithreadEvaluationOption.YES );
> kb = KnowledgeBaseFactory.newKnowledgeBase(kbConf);
> kb.addKnowledgePackages(builder.getKnowledgePackages());
>
> KnowledgeSessionConfiguration sConf =
> KnowledgeBaseFactory.newKnowledgeSessionConfiguration();
> sConf.setOption(ClockTypeOption.get("pseudo"));
> Environment env = EnvironmentFactory.newEnvironment();
> s = kb.newStatefulKnowledgeSession(sConf ,env);
>
> clock = (SessionPseudoClock) s.getSessionClock();
>
> backupStream = s.getWorkingMemoryEntryPoint("Backup Stream");
> }
>
> followed by a test method that looks like this:
>
> public void testLargeBatches() {
>
> Node node1 = new Node("node 1");
> node1.setId(1L);
> s.insert(node1);
> s.fireAllRules();
>
> clock.advanceTime(1L, TimeUnit.DAYS);
> s.fireAllRules();
>
> int batchSize = 40000;
> int daysToRun = 1;
> ArrayList<Backup> batch = new ArrayList<Backup>(batchSize);
> ArrayList<Long> nodeIds = new ArrayList<Long>();
> nodeIds.add(1L);
> DateTime engineTime;
>
> long start;
> long accumulated = 0;
>
> System.err.println("days: "+daysToRun+". batch: "+batchSize+".
> rules: "+ruleFile);
>
> for (int i=0; i<daysToRun; i++) {
> engineTime = new DateTime(clock.getCurrentTime(),
> DateTimeZone.UTC);
> batch.clear();
> //generate random
> BackupGenerationUtil.fillBackupBatch(batch, nodeIds,
> engineTime.minusHours(23).minusMinutes(59).minusSeconds(59),
> engineTime.minusSeconds(1), batchSize);
>
> start = System.currentTimeMillis();
>
> //feed the batch in
> for (Backup b : batch) {
> backupStream.insert(b);
> }
>
> //advance a day and run all rules
> s.fireAllRules();
> clock.advanceTime(1L, TimeUnit.DAYS);
> s.fireAllRules();
>
> accumulated += (System.currentTimeMillis() - start);
> }
>
> long averageBatchProcessingTime =
> (long)(accumulated/(double)daysToRun);
> double perBackupEventTime =
> ((double)accumulated)/(batchSize*daysToRun);
> int eventsPerSecond = (int)(1000D/perBackupEventTime);
>
> System.err.println("average batch processing time is
> "+averageBatchProcessingTime+" which is "+perBackupEventTime+"/event or
> "+eventsPerSecond+" events/second");
> }
>
> running this with or without the 2nd "trimming" rules makes a huge
> difference. on my machine i can run the non-trimming version in batches of
> 300K @~2500 events/second @1GB RAM. with the trimming rule i cant get
> anywhere near that - the above value (40K) is already enough to slow it down
> to ~1800 events/second, and i dont understand whats so "heavy" about my
> attempt to save memory.
>
> any help/thoughts/clues would be most welcome.
>
>
> 2010/12/23 Mauricio Salatino <salaboy at gmail.com>
>
>> hehe yes.. the same happens to "me" :)
>>
>> 2010/12/23 Wolfgang Laun <wolfgang.laun at gmail.com>
>>
>> (OT: your nick "me" is confusing other gmail users. I see your mail as if
>>> I had sent it to myself. I was almost binning it.)
>>>
>>> The delay you see may depend on the way you run the test with 40K events.
>>> Please describe the sequence of insertions and calls to fire...(), the
>>> session setup (pseudo-clock, etc.)
>>>
>>> -W
>>>
>>>
>>> 2010/12/23 me <radai.rosenblatt at gmail.com>
>>>
>>>> Hi.
>>>>
>>>> im trying to demo drools-fusion for a system that processes backup
>>>> events.
>>>> i have the following 2 CEP rules:
>>>>
>>>> rule "Backup Not Succeeded For At Least 3 Days"
>>>> @ruleId(1)
>>>> when
>>>> Node($id : id)
>>>> not ( Backup(clientId == $id, $state: state ==
>>>> BackupStateEnum.FINISHED) over window:time( 3d ) from entry-point "Backup
>>>> Stream" )
>>>> then
>>>> //nothing for now
>>>> end
>>>>
>>>> rule "Prune Previous Successful Backups"
>>>> @ruleId(2)
>>>> when
>>>> $prevBackup : Backup($id : clientId, state ==
>>>> BackupStateEnum.FINISHED) from entry-point "Backup Stream"
>>>> $newerBackup : Backup(clientId == $id, state ==
>>>> BackupStateEnum.FINISHED, this after $prevBackup) from entry-point "Backup
>>>> Stream"
>>>> then
>>>> drools.retract($prevBackup);
>>>> end
>>>>
>>>> rule #2 is my attempt to cut down on memory usage. the problem is that
>>>> rule #2 slows down the processing very very considerably (2 orders of
>>>> magnitude slower when simulating batches of 40K events/day).
>>>>
>>>> what would be my next step in trying to find the reason for it? my
>>>> intuition tells me that with 2 events in the system at any given time
>>>> instead of 120K (single node, 40K events/day 3 day window) i should have
>>>> seen either a speedup or reduced memory consumption (if not both) and yet
>>>> the results im seeing are the exact opposite.
>>>> is there anything im missing ? i could try adding some sort of
>>>> LastSuccessfulBackup "marker" object into main memory and updating it using
>>>> events from the backup stream, but then how would i express rule #1 ?
>>>>
>>>> any help/clues/suggestions would be greatly appreciated,
>>>>
>>>> radai.
>>>>
>>>> _______________________________________________
>>>> rules-users mailing list
>>>> 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
>>>
>>>
>>
>>
>> --
>> - CTO @ http://www.plugtree.com
>> - MyJourney @ http://salaboy.wordpress.com
>> - Co-Founder @ http://www.jbug.com.ar
>>
>> - Salatino "Salaboy" Mauricio -
>>
>> _______________________________________________
>> rules-users mailing list
>> 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/20101226/351b79c0/attachment.html
More information about the rules-users
mailing list