[JDEV] jabber tests - jabber leaks....

[kuba troszok]

Hello,
lately me and my coworker ( Dan Anghel )
have done some tests of the jabber server + jpolld as this is
the installation that is commonly used for
bigger jabber installations and we found some
memory leaks problems.
We were using slightly modified version of
the jabbertestsuite from sf.net.
We modified this test suite because it doesn't
represent exactly the real situation. In
original version it was sending only messages
while there are more possible types of xml tags passing
trough jabber server.

These are the tests:

1. 100 users 2 messages/sec 100 seconds ; 50 users sending, 50 receiving;
   without JPOLLD


2. 100 users 2messages+2presences/sec 100 seconds; 50 users sending, 50
   receiving; without JPOLLD


3. 100 users 2 messages/sec 100 seconds; 50 users sending, 50 receiving; with JPOLLD


4. 100 users 2 messages+2presence/sec 100 seconds; 50 users sending, 50
   receiving; with JPOLLD


Every person had 10 persons in roster so one presence sent
implies 10 <presence> passed trough the server

The problem is that jabberd after tests 1,2,3 is not growing very
much while after test 4 is growing 20 times!!! From 1MB to 22MB!!!
The other problem also is that we have 22% of loss in messages/presence!

Ok the possible source of problem could be jpolld that is not
working as it should work but it is not acceptable why jabberd is growing ?

We found some time ago ( correct me if i'm wrong ) that some packets
are discarded during routing inside jabberd after certain time.
Is it possible that discraded doesn't mean deleted from memory ?
Last week somebody on jdev sent an email explaining the
problem with packet sent to update.jabber.org on every login.
I looks like we have a much more problems like this.

Anyway don't know what packets ( message/presence )
are missing. I'm starting working  on this and i
will try to inform you ASAP.

For a proposition of a how to modify jpolld look at  
next e-mail on jdev at j.o

=====================================================
And here are results:
---------------------------------

TEST 1
MESSAGES STATISTICS
UserPs Rate  Duration MinDelTime MaxDelTime AvgDelTime MsgCnt ExpMsg MsgLossRate
50     2     100      0.00184    2.68925    0.03644    10000  10000  0.00000


Jabberd Memory before:
SIZE 
RSS 
SHARE
1364 
1364 
1036

Jabberd Memory After:

SIZE	RSS	SHARE
1520   1520	1040


slight change

----------------------------------
Test 2:

MESSAGES STATISTICS
UserPs Rate  Duration MinDelTime MaxDelTime AvgDelTime MsgCnt ExpMsg MsgLossRate
50     2     100      0.00375    8.94612    0.15460    10000  10000  0.00000


still no lost messages, but the average delivery time 5 times bigger than
without presence, the max delivery time 4 times bigger.


MEMORY STATISTICS

before:

SIZE	RSS	SHARE
1364	1364	1036

after:

SIZE	RSS	SHARE
1532	1532	1040


everything ok

-----------------------------------------

TEST 3
MESSAGES STATISTICS
UserPs Rate  Duration MinDelTime MaxDelTime AvgDelTime MsgCnt ExpMsg MsgLossRate
50     2     100      0.00279    0.09399    0.01346    10000  10000  0.00000


everything is okay, no messages lost, maybe the average delivery time increased
comparing with jabber alone, without jpolld.


MEMORY STATISTICS

jabberd before:

SIZE	RSS	SHARE 
1340	1340	1012
jabberd after:
SIZE	RSS	SHARE
1520	1520	1024


jpolld before:

RSS	SHARE	STAT
624	624	524 


jpolld after:

SIZE  RSS SHARE 
1352 1352   524       


-----------------------------------------

TEST 4


UserPs Rate  Duration MinDelTime MaxDelTime AvgDelTime MsgCnt ExpMsg MsgLossRate
50     2     100      0.00467    28.27364   8.58076    7741   10000  0.22590


22,59% of messages lost


> MEMORY STATISTICS

jabberd before test:

SIZE	RSS	SHARE
1332	1332	1012



jabberd after test:

SIZE	RSS	SHARE
21340	20M	1024   !!!!!!!!!!!!!!!!!!!!!!!!!!!!

Jpolld before the test
SIZE	RSS	SHARE
624    	624   	524
jpolld after test:
SIZE  	RSS	SHARE 
1416  	1416	524


-------------------------------------------
---kuba troszok
troszok at medialogic.it