uima-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Jaroslaw Cwiklik <uim...@gmail.com>
Subject Re: Unbalanced instances Unexpected behavior for the scaleout parameter
Date Wed, 29 Feb 2012 18:33:31 GMT
Florin, uima-as 2.4.0 has not yet been released. We are still working on
it. Cant give you
a release date at this time.

Jerry

On Wed, Feb 29, 2012 at 10:19 AM, Spico Florin <spicoflorin@gmail.com>wrote:

> Hello!
>  Thank you for your advices. I've found that one of my tasks (Tokenizer,
> the first one from the flow) was responsible for the strange behavior ( one
> very busy thread when the other were free).
> By the way, how I can upgrade to the version 2.4.0 mentioned
> https://issues.apache.org/jira/browse/UIMA-2354? Do you have release
> package or maven repository version?
>
> Thank you.
>  Regards,
>  Florin
>
>  On Wed, Feb 22, 2012 at 8:24 PM, Jaroslaw Cwiklik <uimaee@gmail.com>
> wrote:
>
> > I am not certain, but perhaps you are running into a similar bug reported
> > and fixed recently under
> >
> > https://issues.apache.org/jira/browse/UIMA-2354
> >
> > Maybe the UIMA AS primitive service suffers from a similar problem than
> the
> > UIMA AS aggregate as described in the JIRA. I will verify this soon.
> >
> > Another possibility is that the producer (UIMA AS Aggregate) is just
> > sending enough work to keep a single thread busy in the remote. Do you
> have
> > a CM and large enough cas pool to keep the service busy? You can use
> > jConsole to check the service queue. Look at queue depth and dequeue
> counts
> > while the service is processing. When you colocate the consumer with the
> > rest of the pipeline perhaps the other components contribute to the
> higher
> >  CPU utilization.
> >
> > Jerry C
> >
> > On Wed, Feb 22, 2012 at 10:54 AM, Spico Florin <spicoflorin@gmail.com
> > >wrote:
> >
> > > Hello!
> > >   I have an UIMA pipeline that it is using the
> > > ParallelFlowController(PFC). In ths flow I'm declaring some delegates
> AE
> > > will run  in parallel and serial one(consumer) that that will consume
> the
> > > messages from the first one.
> > >  In the deployment descriptor the consumer is declared as a remote AE.
> > I've
> > > set CASPoolSize of the AAE to 10 and the scaleout parameter to 10.
> > > Profiling the application threads,I've observed that from the 10
> threads
> > > just one of them is very busy doing the job(running mode) and the
> others
> > > most of them time are in the wait mode. Therefore, the bunch of the
> > threads
> > > are unused most of the time. Thus the process time is increasing and
> the
> > > hardware remains unused ( 5% from 2 quad core processors). If I set up
> > the
> > > consumer to be in the same JVM like the other annotators then I have
> the
> > > desired behavior (my processors will be used at 70% and all of the
> > threads
> > > from the pool will process the message). Can you please advice how to
> > > achieve the same behavior with the remote consumer  as having the
> > consumer
> > > locally?
> > > I look forward for your answers and suggestions.
> > >
> > > Thank you.
> > >  Regards,
> > >  Florin
> > >
> > > PFC descriptor descriptor:
> > >
> > > <delegateAnalysisEngine key="OAnnotator">
> > >      <import location="ODesc.xml" />
> > >    </delegateAnalysisEngine>
> > >
> > >    <delegateAnalysisEngine key="GAnnotator">
> > >      <import location="GDesc.xml" />
> > >    </delegateAnalysisEngine>
> > >
> > >    <delegateAnalysisEngine key="RConsumer">
> > >      <import location="RConsumer.xml" />
> > >    </delegateAnalysisEngine>
> > >
> > > <nameValuePair>
> > >        <name>Flow</name>
> > >        <value>
> > >          <array>
> > >            <string>Tokenizer</string>
> > >            <string>OAnnotator,GAnnotator
> > >            </string>
> > >            <string>RConsumer</string>
> > >          </array>
> > >        </value>
> > >      </nameValuePair>
> > >
> >
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message