Eclipse Community Forums
Forum Search:

Search      Help    Register    Login    Home
Home » Modeling » GMT (Generative Modeling Technologies) » Scalability of eCore / oAW
Scalability of eCore / oAW [message #381239] Tue, 11 March 2008 15:22 Go to next message
Eclipse User
Originally posted by: niko.stotz.de.ibm.com

Hi,

Are there any references or some experience with large-scale eCore models?
How would the whole infrastructure react on importing, lets say, a Java
application of two million lines of code (assuming we have a proper
importer)?

Any answer or hint is highly appreciated!

Thanks,

Niko
Re: Scalability of eCore / oAW [message #381759 is a reply to message #381239] Thu, 27 March 2008 14:05 Go to previous messageGo to next message
Sven Efftinge is currently offline Sven Efftinge
Messages: 1762
Registered: July 2009
Senior Member
Hi Niko,

I don't really know, but it's obvious that the additional semantics
(like bidirectional references) of an Ecore model do have an impact on
performance. Also oAW doesn't perform like Java does, because it is
interpreted and it uses dynamic dispatch (i.e. where java binds
everything at compile time, Xpand dispatches at runtime each time a
method/definition/function is being invoked).
So don't expect it to be as fast as an optimized Java solution but
expect it to be easier being developed and maintained ;-)

regards,
Sven

Niko Stotz wrote:
> Hi,
>
> Are there any references or some experience with large-scale eCore models?
> How would the whole infrastructure react on importing, lets say, a Java
> application of two million lines of code (assuming we have a proper
> importer)?
>
> Any answer or hint is highly appreciated!
>
> Thanks,
>
> Niko
>
>


--
Need professional support on Xtext or Xtend?
Mail to: xtext (at) itemis.com
Twitter : @svenefftinge
Blog : blog.efftinge.de
Re: Scalability of eCore / oAW [message #381863 is a reply to message #381759] Thu, 10 April 2008 19:14 Go to previous messageGo to next message
Karsten Thoms is currently offline Karsten Thoms
Messages: 354
Registered: July 2009
Location: Dortmund
Senior Member

Hi Niko!

On the Eclipse DemoCamp in Dortmund Dr. Simon Helsen from SAP told that
EMF does not perform well in ultra-large scale models. Hey, they have
requirements for multi-million elements ;-)

I guess your interest is smaller. And think about alternatives. Are
there any that would perform better than EMF in such scale? Hard to say,
I don't know of any comparisons.

From my experience I would say that EMF performs well for at least
medium-sized models (I'm using UML2 in general).

~Karsten

Sven Efftinge schrieb:
> Hi Niko,
>
> I don't really know, but it's obvious that the additional semantics
> (like bidirectional references) of an Ecore model do have an impact on
> performance. Also oAW doesn't perform like Java does, because it is
> interpreted and it uses dynamic dispatch (i.e. where java binds
> everything at compile time, Xpand dispatches at runtime each time a
> method/definition/function is being invoked).
> So don't expect it to be as fast as an optimized Java solution but
> expect it to be easier being developed and maintained ;-)
>
> regards,
> Sven
>
> Niko Stotz wrote:
>> Hi,
>>
>> Are there any references or some experience with large-scale eCore
>> models? How would the whole infrastructure react on importing, lets
>> say, a Java application of two million lines of code (assuming we have
>> a proper importer)?
>>
>> Any answer or hint is highly appreciated!
>>
>> Thanks,
>>
>> Niko
>>


Need professional support for Xtext, Xpand, EMF?
Go to: http://xtext.itemis.com
Twitter : @kthoms
Blog : www.karsten-thoms.de
Re: Scalability of eCore / oAW [message #381868 is a reply to message #381863] Thu, 10 April 2008 19:30 Go to previous message
Ed Merks is currently offline Ed Merks
Messages: 26054
Registered: July 2009
Senior Member
Karsten,

Comments below.


Karsten Thoms wrote:
> Hi Niko!
>
> On the Eclipse DemoCamp in Dortmund Dr. Simon Helsen from SAP told
> that EMF does not perform well in ultra-large scale models. Hey, they
> have requirements for multi-million elements ;-)
I've heard that Java's heap can't be more that 1.5G so that Java doesn't
scale to really large data sets. :-P
>
> I guess your interest is smaller. And think about alternatives. Are
> there any that would perform better than EMF in such scale? Hard to
> say, I don't know of any comparisons.
There are quite a few things one could do to make EMF instances smaller
(but slower) and there are things like InternalEObject.EStore that could
be used to cache data in some fancy ways.
>
> From my experience I would say that EMF performs well for at least
> medium-sized models (I'm using UML2 in general).
I've certainly spent a lot of time trying to make things fast as well as
reasonably compact. I know folks have done measurements that indicate
that a DOM representation uses up more space and an EMF representation
of the same serialization. I was hoping to work with SAP on a "big
models" project but they have internal technology they are pursuing
instead. I think repositories could potentially help with such type of
scaling, but I think the applications need to change their assumptions
about how to work with models to take the repository into account. I
refer to a repository approach as a closed world model, because the data
space is completely bounded and controlled, whereas EMF's design has
been more of a REST-style open world model, where there are no bounds on
the data space (since a new web page might show up tomorrow with yet
another instance or might exist without us knowing the location).
>
> ~Karsten
>
> Sven Efftinge schrieb:
>> Hi Niko,
>>
>> I don't really know, but it's obvious that the additional semantics
>> (like bidirectional references) of an Ecore model do have an impact
>> on performance. Also oAW doesn't perform like Java does, because it
>> is interpreted and it uses dynamic dispatch (i.e. where java binds
>> everything at compile time, Xpand dispatches at runtime each time a
>> method/definition/function is being invoked).
>> So don't expect it to be as fast as an optimized Java solution but
>> expect it to be easier being developed and maintained ;-)
>>
>> regards,
>> Sven
>>
>> Niko Stotz wrote:
>>> Hi,
>>>
>>> Are there any references or some experience with large-scale eCore
>>> models? How would the whole infrastructure react on importing, lets
>>> say, a Java application of two million lines of code (assuming we
>>> have a proper importer)?
>>>
>>> Any answer or hint is highly appreciated!
>>>
>>> Thanks,
>>>
>>> Niko
>>>
Re: Scalability of eCore / oAW [message #614502 is a reply to message #381239] Thu, 27 March 2008 14:05 Go to previous message
Sven Efftinge is currently offline Sven Efftinge
Messages: 1762
Registered: July 2009
Senior Member
Hi Niko,

I don't really know, but it's obvious that the additional semantics
(like bidirectional references) of an Ecore model do have an impact on
performance. Also oAW doesn't perform like Java does, because it is
interpreted and it uses dynamic dispatch (i.e. where java binds
everything at compile time, Xpand dispatches at runtime each time a
method/definition/function is being invoked).
So don't expect it to be as fast as an optimized Java solution but
expect it to be easier being developed and maintained ;-)

regards,
Sven

Niko Stotz wrote:
> Hi,
>
> Are there any references or some experience with large-scale eCore models?
> How would the whole infrastructure react on importing, lets say, a Java
> application of two million lines of code (assuming we have a proper
> importer)?
>
> Any answer or hint is highly appreciated!
>
> Thanks,
>
> Niko
>
>


--
Need professional support on Xtext or Xtend?
Mail to: xtext (at) itemis.com
Twitter : @svenefftinge
Blog : blog.efftinge.de
Re: Scalability of eCore / oAW [message #614868 is a reply to message #381759] Thu, 10 April 2008 19:14 Go to previous message
Karsten Thoms is currently offline Karsten Thoms
Messages: 354
Registered: July 2009
Location: Dortmund
Senior Member

Hi Niko!

On the Eclipse DemoCamp in Dortmund Dr. Simon Helsen from SAP told that
EMF does not perform well in ultra-large scale models. Hey, they have
requirements for multi-million elements ;-)

I guess your interest is smaller. And think about alternatives. Are
there any that would perform better than EMF in such scale? Hard to say,
I don't know of any comparisons.

From my experience I would say that EMF performs well for at least
medium-sized models (I'm using UML2 in general).

~Karsten

Sven Efftinge schrieb:
> Hi Niko,
>
> I don't really know, but it's obvious that the additional semantics
> (like bidirectional references) of an Ecore model do have an impact on
> performance. Also oAW doesn't perform like Java does, because it is
> interpreted and it uses dynamic dispatch (i.e. where java binds
> everything at compile time, Xpand dispatches at runtime each time a
> method/definition/function is being invoked).
> So don't expect it to be as fast as an optimized Java solution but
> expect it to be easier being developed and maintained ;-)
>
> regards,
> Sven
>
> Niko Stotz wrote:
>> Hi,
>>
>> Are there any references or some experience with large-scale eCore
>> models? How would the whole infrastructure react on importing, lets
>> say, a Java application of two million lines of code (assuming we have
>> a proper importer)?
>>
>> Any answer or hint is highly appreciated!
>>
>> Thanks,
>>
>> Niko
>>


Need professional support for Xtext, Xpand, EMF?
Go to: http://xtext.itemis.com
Twitter : @kthoms
Blog : www.karsten-thoms.de
Re: Scalability of eCore / oAW [message #614872 is a reply to message #381863] Thu, 10 April 2008 19:30 Go to previous message
Ed Merks is currently offline Ed Merks
Messages: 26054
Registered: July 2009
Senior Member
Karsten,

Comments below.


Karsten Thoms wrote:
> Hi Niko!
>
> On the Eclipse DemoCamp in Dortmund Dr. Simon Helsen from SAP told
> that EMF does not perform well in ultra-large scale models. Hey, they
> have requirements for multi-million elements ;-)
I've heard that Java's heap can't be more that 1.5G so that Java doesn't
scale to really large data sets. :-P
>
> I guess your interest is smaller. And think about alternatives. Are
> there any that would perform better than EMF in such scale? Hard to
> say, I don't know of any comparisons.
There are quite a few things one could do to make EMF instances smaller
(but slower) and there are things like InternalEObject.EStore that could
be used to cache data in some fancy ways.
>
> From my experience I would say that EMF performs well for at least
> medium-sized models (I'm using UML2 in general).
I've certainly spent a lot of time trying to make things fast as well as
reasonably compact. I know folks have done measurements that indicate
that a DOM representation uses up more space and an EMF representation
of the same serialization. I was hoping to work with SAP on a "big
models" project but they have internal technology they are pursuing
instead. I think repositories could potentially help with such type of
scaling, but I think the applications need to change their assumptions
about how to work with models to take the repository into account. I
refer to a repository approach as a closed world model, because the data
space is completely bounded and controlled, whereas EMF's design has
been more of a REST-style open world model, where there are no bounds on
the data space (since a new web page might show up tomorrow with yet
another instance or might exist without us knowing the location).
>
> ~Karsten
>
> Sven Efftinge schrieb:
>> Hi Niko,
>>
>> I don't really know, but it's obvious that the additional semantics
>> (like bidirectional references) of an Ecore model do have an impact
>> on performance. Also oAW doesn't perform like Java does, because it
>> is interpreted and it uses dynamic dispatch (i.e. where java binds
>> everything at compile time, Xpand dispatches at runtime each time a
>> method/definition/function is being invoked).
>> So don't expect it to be as fast as an optimized Java solution but
>> expect it to be easier being developed and maintained ;-)
>>
>> regards,
>> Sven
>>
>> Niko Stotz wrote:
>>> Hi,
>>>
>>> Are there any references or some experience with large-scale eCore
>>> models? How would the whole infrastructure react on importing, lets
>>> say, a Java application of two million lines of code (assuming we
>>> have a proper importer)?
>>>
>>> Any answer or hint is highly appreciated!
>>>
>>> Thanks,
>>>
>>> Niko
>>>
Previous Topic:Re: [AM3][ATL] Experiences getting ATL/AM3/TCS running from CVS
Next Topic:How to install KM3 to Eclipse 3.3.2
Goto Forum:
  


Current Time: Sun Sep 21 16:17:48 GMT 2014

Powered by FUDForum. Page generated in 0.04510 seconds
.:: Contact :: Home ::.

Powered by: FUDforum 3.0.2.
Copyright ©2001-2010 FUDforum Bulletin Board Software