Overhead in running UIMA Asynchronous Scaleout

[ajay kumar <aj...@gmail.com>]

Hello,
I am trying to compare the runtimes for pipelines running locally and
pipelines using UIMA AS.
I have a setup of each annotators in a pipeline running on different cores.
It shows from the results pipelines in UIMA AS takes around 4seconds to
complete processing which is a huge delay compared to local pipelines. I
wanted to know how much delay is induced from the flow controller
forwarding each CAS and the reply queues.

Re: Overhead in running UIMA Asynchronous Scaleout

[Jaroslaw Cwiklik <cw...@apache.org>]

Do you deploy both the uima-as client and a service in the same process?
How do you get your timing? There is a CAS serialization/deserialization
penalty and latency in a broker but should not be huge. The design goal is
to support scaleout of analytics across multiple processes on multiple
machines. For maximum performance we recommend sending a CAS with a
reference to data to be analyzed instead of sending data in the CAS.

How big is the CAS you are sending from the client? If you are using a
default FC, there should be no significant delay there.

Jerry

On Mon, Jul 31, 2017 at 5:01 AM, ajay kumar <aj...@gmail.com>
wrote:

> Hello,
> I am trying to compare the runtimes for pipelines running locally and
> pipelines using UIMA AS.
> I have a setup of each annotators in a pipeline running on different cores.
> It shows from the results pipelines in UIMA AS takes around 4seconds to
> complete processing which is a huge delay compared to local pipelines. I
> wanted to know how much delay is induced from the flow controller
> forwarding each CAS and the reply queues.
>