incubator-s4-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Matthieu Morel <mmo...@apache.org>
Subject Re: Fault tolerance and communication
Date Fri, 22 Mar 2013 17:50:21 GMT

On Mar 21, 2013, at 11:56 , Dingyu Yang wrote:

> Hi,Matthieu 
> The version is S4 0.6 RC2.
> As I know, if the key is null, every event is send to one node, depending on round robin.
>  I just use the code to send the event to app cluster: getRemoteStream().put(event);
> The adapter is not keyed. The splitPE is singleton as twitter example.
> What do you mean : a given operation is repeated for each target node.
> It doesn't make sense of increasing node, the time increases.

Well it might make sense if you were doing synchronous and duplicate communications, or if
the app nodes need to synchronize with some external service, or for some other reason.


One point we can observe here is that the default RemoteSendersExecutorServiceFactory implementation
is blocking by default, so that the upstream system (adapter) does not send more than what
can be processed by the downstream system.

You can of course change this implementation, by providing your own overriding module, but
with the default blocking behaviour, the pattern you observe seems to indicate some particular
behaviour in your downstream application. 


Hopefully that can help you.


Let us know!

Matthieu

 




> 
> Dingyu
> 
> >
> > Another problem is that the communication between adapter and app.
> > I test the experiment of word count, a 500M file with 80775764 words.
> > multiple nodes for app partitions, one node for adapter.
> > I test one adatper node and one app node, the adapter sending all the words
> > is done with 35 seconds.
> > one adatper node and two app node, the adapter is done with 61 seconds.
> > one adatper node and three app node, the adapter is done with 95 seconds.
> >
> > The adapter node is a same node and same program.
> > The time of adapter should be same or less with increasing app nodes, since
> > its processing ability has increased.
> > I don't know what the problem is.
> 
> There were some extra copies in S4 0.5 so if you are using that version it could be an
explanation.
> 
> The pattern is quite clear though (linear increase with number of nodes) and it should
be easy to spot the issue. Looks like a given operation is repeated for each target node.
Are you broadcasting to all nodes? Are the events from the adapter keyed? Is there something
specifically related in your adapter app code or adapter app graph?
> 
> 
> Regards,
> 
> Matthieu
> 
> 


Mime
View raw message