Just tried the CVS head and everything works with any CPU going crazy! only if ld_assume_kernel is set to 2.4
One more question for you Filip, is the useDirtyFlag working at all? It seams like even if it's set to true, the whole session gets replicated after each request. :(
Jean-Philippe
[EMAIL PROTECTED] wrote:
Hurray for Fillip! :)
I'll get the CVS head for the module today and test this out. Happy to see that it got fixed that quickly!
Thanks again and I'll let you know how it goes
Jean-Philippe
Filip Hanik wrote:
Jean-Philippe and Steve,
I fixed the bug, and tried replication on RH9. Immediately it didn't work.
The problem is that when RH9 tries to write the ACK back to the NIO socket,
it never reaches the other node. and times out after a long time.
I set LD_ASSUME_KERNEL=2.4 and it started to work
Filip
-----Original Message----- From: Filip Hanik [mailto:[EMAIL PROTECTED] Sent: Thursday, January 08, 2004 6:43 PM To: Tomcat Users List Subject: RE: tomcat 5.0.16 Replication
ok guys, good news. The 100% cpu is totally my fault. I messed up on that one. I was registering OP_WRITE as an interest this is not good :) checking in the working code in 15 min, some more regression tests Filip
-----Original Message----- From: Filip Hanik [mailto:[EMAIL PROTECTED] Sent: Thursday, January 08, 2004 2:54 PM To: Tomcat Users List Subject: RE: tomcat 5.0.16 Replication
another code change was, that I am now accepting keys for OP_READ and OP_WRITE. before it was only OP_READ, but for synchronous replication I need both.
this is good info, I just got RH9 installed. will be trying it out this and
next week.
Filip
-----Original Message----- From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED] Sent: Thursday, January 08, 2004 11:46 AM To: Tomcat Users List Subject: Re: tomcat 5.0.16 Replication
The only changes in the ReplicationListener class is the try catch that was added.
the code logic is the same. Weird enough. So it's probably elsewhere that something changed in the state of the SelectionKey.
Jean-Philippe Bélanger
Steve Nelson wrote:
I was just about to try this actually. I found through googling alot of people having problems with select with 1.4 and NIO with Redhat 9. They were actually experiencing crashes though.
To verify your results I just put a Thread.Sleep(1); where you
suggested and
I also see the jump in performance.
Something must have changed in ReplicationListener that causes this because
the 5.0.16
version doesn't seem to have the problem. I'll see if I can figure it out
when I get back to where I can diff the files.
-Steve
-----Original Message----- From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED] Sent: Thursday, January 08, 2004 12:25 PM To: Tomcat Users List Subject: Re: tomcat 5.0.16 Replication
More content for you Filip.
I've checked and followed the code of the listen event in ReplicationListener.java
Here's what happening:
selector.select(timeout) -> return immediatly with one SelectorKey
available
That key is not Acceptable and not Readable so it immediatly skip those IFs and loops back to the beginning.
I've put traces and this is executed once every millisecond hence the 100% load on the server. Just to make sure, I've put a Thread.sleep(10) at the end of the loop and the CPU dropped back to 0% and the replication still worked nicely but probably a little slower since the wait of 10ms.
I don't know much about those NIO packages but seams like the select(timeout) method shouldn't return a SelectorKey of that state. with any waiting.
Let me know what you can dig from those.
Jean-Philippe Bélanger
[EMAIL PROTECTED] wrote:
Hi Filip.
I did some profiling of 40mins of tomcat with and without a 2nd node
up. here are the results with
-Xrunhprof:cpu=samples,thread=y,file=/u01/portal/java.hprof.txt,depth=10:
Those number are cpu=times and not samples since the later one freezes on my systems. So that list shows the time spent in each methods.
Major difference the some call to the sun.nio.ch.PollArrayWrapper class. I don't know much about those NIOs packages but 819000 call in 40 mins is a lot. The Socket Interface was called more than twice with 2 hosts than with a single one. Which seams normal.
Maybe this can help. If you need the complete hprof file I can send them to you.
1 host in cluster:
CPU TIME (ms) BEGIN (total = 19701) Thu Jan 8 10:00:59 2004
rank self accum count trace method
1 11.48% 11.48% 54 85 java.lang.Object.wait
2 11.46% 22.94% 117 86 java.lang.Object.wait
3 10.95% 33.89% 4115 215 java.net.PlainDatagramSocketImpl.receive
4 10.93% 44.81% 4114 224 java.lang.Thread.sleep
5 10.91% 55.73% 19005 214 sun.nio.ch.PollArrayWrapper.poll0
6 7.37% 63.09% 28 495 java.lang.Object.wait
7 7.24% 70.34% 10 576 java.lang.Object.wait
8 4.57% 74.90% 90 716 java.lang.Thread.sleep
9 4.48% 79.38% 1 909 java.lang.Object.wait
10 4.48% 83.86% 1 908 java.lang.Object.wait
11 4.48% 88.34% 15 810 java.lang.Object.wait
12 4.47% 92.81% 1 910 java.net.PlainSocketImpl.socketAccept
13 0.71% 93.52% 2 623 java.lang.Object.wait
14 0.56% 94.08% 2 706 java.lang.Object.wait
15 0.38% 94.46% 2 914 java.lang.Object.wait
16 0.24% 94.70% 775 913 java.lang.String.toCharArray
17 0.23% 94.93% 3 475 java.lang.Thread.sleep
18 0.16% 95.09% 2 472 java.lang.Object.wait
19 0.15% 95.24% 2 595 java.lang.Thread.sleep
20 0.15% 95.40% 2 586 java.lang.Thread.sleep
21 0.15% 95.55% 2 703 java.lang.Thread.sleep
22 0.15% 95.70% 2 476 java.lang.Thread.sleep
23 0.15% 95.85% 2 692 java.lang.Thread.sleep
24 0.12% 95.97% 218595 385 java.lang.CharacterDataLatin1.toLowerCase
25 0.12% 96.09% 218595 408 java.lang.Character.toLowerCase
26 0.11% 96.20% 218595 433
java.lang.CharacterDataLatin1.getProperties
27 0.10% 96.30% 210925 389 java.lang.String.equalsIgnoreCase
28 0.08% 96.38% 157259 387 java.lang.String.charAt
29 0.08% 96.46% 1 646 java.lang.Thread.sleep
30 0.08% 96.53% 1 634 java.lang.Thread.sleep
31 0.08% 96.61% 1 903 java.lang.Thread.sleep
32 0.08% 96.69% 1 714 java.lang.Thread.sleep
33 0.08% 96.76% 1 811 java.lang.Thread.sleep
34 0.08% 96.84% 1 715 java.lang.Thread.sleep
2 hosts: CPU TIME (ms) BEGIN (total = 37247) Thu Jan 8 11:01:28 2004 rank self accum count trace method 1 9.56% 9.56% 52 85 java.lang.Object.wait 2 9.56% 19.12% 29 86 java.lang.Object.wait 3 9.30% 28.43% 3 267 java.lang.Object.wait 4 9.25% 37.68% 6644 224 java.lang.Thread.sleep 5 9.23% 46.91% 13116 215 java.net.PlainDatagramSocketImpl.receive 6 7.67% 54.58% 3 266 java.lang.Object.wait 7 5.90% 60.47% 39 847 java.lang.Object.wait 8 5.76% 66.24% 12 503 java.lang.Object.wait 9 3.90% 70.14% 145 975 java.lang.Thread.sleep 10 3.90% 74.04% 1 1174 java.lang.Object.wait 11 3.90% 77.94% 1 1173 java.lang.Object.wait 12 3.90% 81.84% 25 973 java.lang.Object.wait 13 3.90% 85.74% 1 1175 java.net.PlainSocketImpl.socketAccept 14 3.88% 89.62% 819692 214 sun.nio.ch.PollArrayWrapper.poll0 15 0.75% 90.37% 2 958 java.lang.Object.wait 16 0.28% 90.65% 2 457 java.lang.Object.wait 17 0.26% 90.91% 2 1181 java.lang.Object.wait
Filip Hanik wrote:
I'll try to get an instance going today. Will let you know how it goes
also, try asynchronous replication, does it still go to 100%?
Filip
-----Original Message----- From: Steve Nelson [mailto:[EMAIL PROTECTED] Sent: Wednesday, January 07, 2004 12:08 PM To: 'Tomcat Users List' Subject: RE: tomcat 5.0.16 Replication
Okay, did that got this
BEGIN TO RECEIVE SENT:Default 1 RECEIVED:Default 1 FROM /10.0.0.110:5555 SENT:Default 2 BEGIN TO RECEIVE RECEIVED:Default 2 FROM /10.0.0.110:5555 SENT:Default 3 BEGIN TO RECEIVE RECEIVED:Default 3 FROM /10.0.0.110:5555 SENT:Default 4 BEGIN TO RECEIVE RECEIVED:Default 4 FROM /10.0.0.110:5555
*shrug*
BTW It didn't go to 100% CPU ute before I started using the code from CVS. Of course the Manager would almost always timeout before it would recieve the message.
Now it gets the message right away, but maxes my machine out.
-----Original Message----- From: Filip Hanik [mailto:[EMAIL PROTECTED] Sent: Wednesday, January 07, 2004 1:58 PM To: Tomcat Users List Subject: RE: tomcat 5.0.16 Replication
100% cpu can mean that you have a multicast problem, try to run
java -cp tomcat-replication.jar MCaster
download the jar from http://cvs.apache.org/~fhanik/
Filip
-----Original Message----- From: Steve Nelson [mailto:[EMAIL PROTECTED] Sent: Wednesday, January 07, 2004 6:51 AM To: '[EMAIL PROTECTED]' Subject: tomcat 5.0.16 Replication
I was having random problems with clustering when starting up. Mostly it had to do with Timing out when the manager was starting up. I built the CVS version and it solved that problem. But it has caused some serious performance problems.
First a little background.
I have 2 servers, dual 300mhz cpq proliants, both running Redhat - 9, Tomcat 5.0.16 (with catalina-cluster.jar build from cvs) The multicast packets are restricted to a crossover link between the servers. There are 3 hosts in the server.xml, all with clustering set up. They all function just fine.
But.....the cpu's spikes up to 100% if I start up both servers. I
know this
didn't happen without the new catalina-cluster.jar. If I shut down 1
server
(doesn't matter which) everything returns to normal. But when both are
running both servers are at 100% CPU. I am trying to profile it now,
but I
figured if someone has already experienced this they could save me some
time.
Oh, and there isn't anything relevant in my logs. It's not throwing millions of errors or something.
-Steve Nelson
--------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
--------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
--------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
-- Jean-Philippe Bélanger (514)228-8800 ext 3060 111 Duke CGI
--------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
--------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
--------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
--------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
-- Jean-Philippe Bélanger (514)228-8800 ext 3060 111 Duke CGI
--------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]