Re: [Oscar-devel] OFED

[Paul G. <pau...@ua...>]

Hi Geoffroy,

I did some performance test with POV-Ray, and I went from 19 minutes to
14 minutes, saving 25% of the runtime.  I am working on doing some
similar comparisons for POP and GAMESS right now as well.

I'm not actually thinking that distributing the Intel compiler (or any
others) is possible.. I need to dig into the licenses and see if we can
distribute OFED compiled with Intel, PGI and whatever else we have
access to.  The final user would have to manually put the Intel compiler
and the libraries in place to be able to use it, but with documentation
and scheduler integration, it's more likely to be used.  Alternatively,
just having the pieces so OSCAR can be pointed at a OFED installation,
detect what's installed, and put the switcher and scheduler integration
into place might be sufficient..

Paul

geo...@fr... wrote:
> Hi Joe,
> 
> Thanks for the details, it is actually very interesting to have some return of
> experience. However, my question was actually very unclear (my bad), I was
> actually looking for more details about the performance gain (10% in average for
> MPI applications?).
> I also know that if you start to play with the compiling options you can improve
> the global performance of your app (both with gcc and the Intel compiler). I am
> just curious, are your users only using the default compiling options? or do
> they try to tune the compiling procedure for each compiler?
> I also wonder if the latest version of GCC will improve performance compared to
> the Intel compiler; it is supposed to include some auto-vectorization
> mechanism). But that's another story, i digress here. :-)
> 
> On the OSCAR side, the problem with the Intel compiler is the license: we cannot
> install automatically the Intel compiler without asking the user to agree with
> Intel licensing stuff, and clearly, currently the only solution to do so is via
> a hack (we do not have a well-defined method for that).
> To summarize, my concern is the following: why should we ship OFED compiled with
> the Intel compiler if you do not have a good support of the Intel compiler by
> default. My feeling is the usage of the Intel compiler will really be a plus if
> we can also compile the application. In other term, it seems to me that instead
> of working on a OFED OPKG compiled with the Intel compiler we should be better
> off working on a OPKG for the Intel compiler.
> Do you agree with that?
> 
> 
> Selon "Greenseid, Joseph M." <Jos...@ng...>:
> 
>> My users find that using the Intel compilers do increase application
>> performance for them over GCC (especially with x86_64 architecture, because
>> of the auto-vectorization you can get with the SSE aware compilers).
>>
>> Given the choice, none of the folks using my system choose GCC over Intel
>> (currently, all my users use Intel compilers and Intel flavors of MPI, even
>> though GCC is also available).
>>
>> I would say that the ability to use the Intel compilers with MPI would be
>> valuable, if OFED is would be considered in OSCAR.  If Intel flavors aren't
>> included, I would probably have to choose to not install it via OSCAR, but
>> instead do it myself by hand to get them.
>>
>> Just the thoughts of one user...
>>
>> --Joe
>>
>> ________________________________
>>
>> From: osc...@li... on behalf of
>> geo...@fr...
>> Sent: Tue 3/25/2008 12:03 AM
>> To: osc...@li...; Paul Greidanus
>> Subject: Re: [Oscar-devel] OFED
>>
>>
>>
>> Hi,
>>
>> Just another idea: instead of focusing on the support of multiple compiler
>> (does
>> it really improve the global performance for all applications?), why not
>> working
>> on the virtualization stuff that are today included in OFED (Panda team work
>> typically), i.e., VMM-bypass and efficient VM migration?
>>
>> That's should be fun to do and it fits perfectly the OSCAR-V extension
>> (which, i
>> hope, will be very soon integrated directly into OSCAR).
>>
>> My 2 cents,
>>
>> Selon Paul Greidanus <pau...@ua...>:
>>
>>> Hi All,
>>>
>>> I was just looking and thinking about OFED, and how to package it into
>>> OSCAR, and I'm thinking it might be possible to just copy the SRPMS in,
>>> and have the oscar build scripts build the RPMS, like the rest of
>>> oscar?  Also, we could use an Intel/pgi compiler to build the RPMS
>>> pretty easily as well, and have the option of installing optimized
>>> libraries, rather then just GCC..
>>>
>>> Or do I not know what I'm talking about..
>>>
>>> Paul
>>>
>>> -------------------------------------------------------------------------
>>> This SF.net email is sponsored by: Microsoft
>>> Defy all challenges. Microsoft(R) Visual Studio 2008.
>>> http://clk.atdmt.com/MRT/go/vse0120000070mrt/direct/01/
>>> _______________________________________________
>>> Oscar-devel mailing list
>>> Osc...@li...
>>> https://lists.sourceforge.net/lists/listinfo/oscar-devel
>>>
>>
>>
>> -------------------------------------------------------------------------
>> This SF.net email is sponsored by: Microsoft
>> Defy all challenges. Microsoft(R) Visual Studio 2008.
>> http://clk.atdmt.com/MRT/go/vse0120000070mrt/direct/01/
>> _______________________________________________
>> Oscar-devel mailing list
>> Osc...@li...
>> https://lists.sourceforge.net/lists/listinfo/oscar-devel
>>
>>
>>
> 
> 
> 
> -------------------------------------------------------------------------
> Check out the new SourceForge.net Marketplace.
> It's the best place to buy or sell services for
> just about anything Open Source.
> http://ad.doubleclick.net/clk;164216239;13503038;w?http://sf.net/marketplace
> _______________________________________________
> Oscar-devel mailing list
> Osc...@li...
> https://lists.sourceforge.net/lists/listinfo/oscar-devel
>