tomcat-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Filip Hanik - Dev Lists <>
Subject Re: DTrace analysis of TC-Cluster Multicast
Date Sat, 25 Feb 2006 23:35:27 GMT
correctness can be achieved in many different ways,  synchronization 
just happens to be the easiest one.
The build is working, including session replication (modules 
groupcom,ha,storeconfig-ha), just uncomment one line in 
container/modules/ and tomcat builds the new modules 
instead of the old.
I have a lot of refactoring and optimization left to do.


Rainer Jung wrote:
> I agree, heartbeat is not that important w.r.t. performance. I used it 
> as a simple test for how DTrace will be able to help. And at least in 
> that nearly trivial test case the results are already interesting.
> Another thing that DTrace can check is lock contention. So there's 
> more to come :)
> Do you have a working build of groupcom? I could trace some of the 
> group communication.
> I hope you will not sacrifice synchronization correctness for 
> performance ...
> Rainer
> Filip Hanik - Dev Lists wrote:
>> Cool Rainer, interesting numbers. Can you run the same numbers on the 
>> new module groupcom.
>> I have done some optimizations there, for example, the name isn't 
>> broadcasted. And I am about to remove the domain, instead of having a 
>> default one. also, I agree, lets not translate the 4byte IP and the 
>> 4byte port when receiving it. instead, member comparisons can be done 
>> on the byte array level, excellent suggestion.
>> After that I won't be spending to much focus on this issue, as a 
>> heartbeat do not really have to be optimized. The heartbeat runs 
>> *very* infrequently and it not a critical piece.
>> Instead I am gonna start reworking the data senders, there are way 
>> too many synchronized calls locking down the sockets, the code is 
>> somewhat cluttered then I am gonna move on to serialization. In the 
>> current cluster module, too much gets serialized, in the new module 
>> "ha" I have already removed a lot of serialization and will continue 
>> stripping it out.
>> Filip

To unsubscribe, e-mail:
For additional commands, e-mail:

View raw message