Fenderbosch, Eric wrote:
We are having a similar problem, although our fact count is much higher.
Performance seems pretty good and consistent until about 400k facts,
then performance degrades significantly. Part of the degradation is
from bigger and more frequent GCs, but not all of it.
If you have multi-cpu there is a JVM command you can set a dedicated cpu
for GC, that helps somewhat.
Time to load first 100k facts: ~1 min
Time to load next 100k facts: ~1 min
Time to load next 100k facts: ~2 min
Time to load next 100k facts: ~4 min
This trend continues, going from 600k to 700k facts takes over 7
minutes. We're running 4.0.7 on a 4 CPU box with 12 GB, 64 bit RH Linux
and 64 bit JRockit 5. We've allocated a 9 GB heap for the VM using
large pages, so no memory paging is happening. JRockit is started w/
the -XXagressive parameter, which enables large pages and the more
efficient hash function in HashMap which was introduced in Java5 update
8.
Other than the CPU thing, Drools won't take advantage of multipe cpus at
the moment.
http://e-docs.bea.com/jrockit/jrdocs/refman/optionXX.html
The end state is over 700k facts, with the possibility of nearly 1M
facts in production. After end state is reached and we issue a few GC
requests, if looks like our memory per fact is almost 9k, which seems
quite high as most of the facts are very simple. Could that be due to
our liberal use of insertLogical and TMS?
It could be related to this, especially if you create a long chain of
logical relationships.
We've tried performing a "commit" every few hundred fact insertions by
issuing a fireAllRules periodically, and that seems to have helped
marginally.
I tried disabling shadow proxies and a few of our ~390 test cases fail
and one loops indefinitely. I'm pretty sure we could fix those, but
don't want to bother if this isn't a realistic solution.
Any thoughts?
Have you tried this on Drools 5.0? It 'doesn't need shadow proxies and
implements a new Rete algorithm that is faster for retracts. You can get
a nightly build from here, I'd be interested to find out how broken 5.0
is :)
https://hudson.jboss.org/hudson/job/drools/lastSuccessfulBuild/artifact/trunk/target/
We still have more performnace work to do, the items are known, just a
matter of time, not all will make 5.0 though. but the main items include:
1) bytecode compiled Rete network, instead of interpreted nodes. I'm
hoping this will have a large impact, reducing GC and general
indirection and recursive method call frames.
2) "true modify", instead of a retract+assert, will also remove the need
for activation normalistaion that we do for TMS and the agenda event model.
3) range indexing (initially literals, but would like to explore
variables too).
Steve, before he left fedex, was creating a simulator for this use case,
but removing anything business sensitive. So that we could use it
publicly as a benchmark and to help us tune the engine. Are you still
working on this? Steve use to chat to us on irc, can I ask you to pop on
for a chat?
http://labs.jboss.org/drools/irc.html
mark
Thanks
Eric
-----Original Message-----
From: [EMAIL PROTECTED]
[mailto:[EMAIL PROTECTED] On Behalf Of Ron Kneusel
Sent: Thursday, June 26, 2008 12:47 PM
To: rules-users@lists.jboss.org
Subject: [rules-users] Drools 4 poor performance scaling?
I am testing Drools 4 for our application and while sequential mode is
very fast I get very poor scaling when I increase the number of facts
for stateful or stateless sessions. I want to make sure I'm not doing
something foolish before deciding on whether or not to use Drools
because from what I am reading online it should be fast with the number
of facts I have.
The scenario: I have 1000 rules in a DRL file. They are all of the
form:
rule rule0000
when
Data(type == 0, value> 0.185264);
Data(type == 3, value < 0.198202);
then
insert(new AlarmRaised(0));
warnings.setAlarm(0, true);
end
where the ranges checked on the values and the types are randomly
generated. Then, I create a Stateful session and run in a loop timing
how long it takes the engine to fire all rules as the number of inserted
facts increases:
// Run
for(j=0; j < 100; j+=5) {
if (j==0) {
nfacts = 1;
} else {
nfacts = j;
}
System.out.println(nfacts + ":");
// Get a working memory
StatefulSession wm = ruleBase.newStatefulSession();
// Global - output
warnings = new Alarm();
wm.setGlobal("warnings", warnings);
// Add facts
st = (new Date()).getTime();
for(i=0; i < nfacts; i++) {
wm.insert(new Data(rand.nextInt(4),
rand.nextDouble()-0.5));
}
en = (new Date()).getTime();
System.out.println(" facts = " + (en-st));
// Now run the rules
st = (new Date()).getTime();
wm.fireAllRules();
en = (new Date()).getTime();
System.out.println(" rules = " + (en-st));
// Clean up
wm.dispose();
System.out.println("\n");
}
This code is based on the HelloWorldExample.java code from the manual
and the setup for the rule base is the same as in the manual. As the
number of facts increases runtime increases dramatically:
facts -- runtime (ms)
10 -- 168
20 -- 166
30 -- 344
40 -- 587
50 -- 1215
60 -- 1931
70 -- 2262
80 -- 3000
90 -- 4754
with a maximum memory use of about 428 MB RAM. By contrast, if I use
sequential stateless sessions, everything runs in about 1-5 ms.
Is there something in my set up that would cause this, or is this how
one would expect Drools to scale? I read about people using thousands
of facts so I suspect I'm setting something up incorrectly.
Any help appreciated!
Ron
_________________________________________________________________
The other season of giving begins 6/24/08. Check out the i'm Talkathon.
http://www.imtalkathon.com?source=TXT_EML_WLH_SeasonOfGiving
_______________________________________________
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
