ah, events werent being evicted because RHS was only evaluated during fire(), which was after a batch.
i have altered my test code according to your suggestion, to look like this:

public void testLargeBatches() {
        Node node1 = new Node("node 1");
        node1.setId(1L);
        s.insert(node1);

        clock.advanceTime(1L, TimeUnit.DAYS);

        int batchSize = 40000;
        int daysToRun = 10;
        ArrayList<Backup> batch = new ArrayList<Backup>(batchSize);
        ArrayList<Long> nodeIds = new ArrayList<Long>();
        nodeIds.add(1L);
        DateTime engineTime;

        long start;
        long accumulated = 0;
        long currentTime;
        long eventTime;
        long delta;

        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
            System.err.println("preparing batch " + i);
            BackupGenerationUtil.fillBackupBatch(batch, nodeIds, engineTime.minusHours(23).minusMinutes(59).minusSeconds(59), engineTime.minusSeconds(1), batchSize);

            System.err.println("feeding batch "+i);
            start = System.currentTimeMillis();

            currentTime = clock.getCurrentTime();

            //feed the batch in
            for (Backup b : batch) {
                eventTime = b.getFinished().getTime();
                delta = (eventTime-currentTime);
                clock.advanceTime(delta, TimeUnit.MILLISECONDS);
                currentTime = clock.getCurrentTime();
                backupStream.insert(b);
                s.fireAllRules();
            }

            //advance a day (whats left of it)
            delta = engineTime.plusDays(1).getMillis()-currentTime;
            clock.advanceTime(delta, TimeUnit.MILLISECONDS);
            s.fireAllRules();

            accumulated += (System.currentTimeMillis() - start);

            System.err.println("batch complete "+i);
        }

        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");
}

the batch is now ordered by event end time (inside fillBackupBatch() ), the clock is advanced to the end of each event as it is processed and rules are fired after each insertion. this code will eventually be run inside a j2ee environment so i would rather not use fireUntilHalt() (also, using it maxed-out a single core - does it do busy-waiting internally?!). also, the batch nature is by design. the system will process events in a batch fashion, not truly one by one. in production there will be many such batches per day, but they could be very sizable.

for 40K events/day over 10 days, the above code (on my machine) lands me 59K events/sec without the 2nd rule and 5300 events/sec with the second rule. while this is an improvement of x2-x3 compared to my prev. code (24k without the 2nd rule and 1400 with), for which i am very glad, i still cannot understand why the 2nd rule slows me down by x10.

anything else i might do to get the performance if the 2 rules close to the performance of the single rule?

2010/12/26 Wolfgang Laun <wolfgang.laun@gmail.com>
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@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@gmail.com>
hehe yes.. the same happens to "me" :)

2010/12/23 Wolfgang Laun <wolfgang.laun@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@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@lists.jboss.org
https://lists.jboss.org/mailman/listinfo/rules-users



_______________________________________________
rules-users mailing list
rules-users@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@lists.jboss.org
https://lists.jboss.org/mailman/listinfo/rules-users



_______________________________________________
rules-users mailing list
rules-users@lists.jboss.org
https://lists.jboss.org/mailman/listinfo/rules-users



_______________________________________________
rules-users mailing list
rules-users@lists.jboss.org
https://lists.jboss.org/mailman/listinfo/rules-users