Re: [javagroups-users] Xmit Requests

[Bela B. <be...@ya...>]

I still don't understand your distributed queue. If you have a small 
test case, that would help

fungrim wrote:
> Hi Bela,
>
> I'm working with Fredrik on this one. Our man with the hardware skills will
> be back from vacation next week and we'll start investigating the systems in
> general and the switch in particular. 
>
> The setup is a distributed data set (similar to a queue). In the tests node
> 2 was producer, node 3 consumer and node 1 backup. The data set uses local
> (non-distributed) events for addition and removal, meaning that node 2 got
> removal over jgroups only, node 3 additions over jgroups only, but node 1
> got both, ie. twice as much trafic over the jchannel.
>
> Our memory problems was due to out of order packets (or even dropped ones).
> As we reached a certain bandwith (aprox 10 mbit/sec) we'd start drop
> packages abnormally, which would escalate NAKACK xmit requests and cause a
> "leak" in our distributed data set.
>
>
> Bela Ban wrote:
>   
>> That's entirely possible, as some switches don't handle IP multicast 
>> packets well. Note that in my tests, I used a GB switch, so that at 
>> least explains the diff between your and my numbers. However, despite 
>> the switch, you should not run out of memory.
>> [...]
>> Don't understand this: is everyone a sender ?
>>
>>     
>
>   

-- 
Bela Ban
Lead JGroups / JBoss Clustering team
JBoss - a division of Red Hat