[Q-e-developers] Questions About Optimization and Parallelism of Quantum Espresso

Thu Oct 20 06:58:43 CEST 2016

Dear all, 
  I'm working on a different R&G distribution that could scale more favorably w number of processors. More will be presenter at next January developers meeting. It may appears in the distribution even before then if I'm happy with it.
I'm in contact with CarloC and PaoloG about that,
Best
stefano 
(sent from my phone)

> On 20 Oct 2016, at 02:44, Ye Luo <xw111luoye at gmail.com> wrote:
> 
> Dear Carlo,
> 
> It is very enlightening to read your comment on architecture and QE.
> Do you have more recent technical talks about the re-factoring of QE?
> My knowledge is still limited to your talks during 2012 when BG/Q is introduced and QE dev meetings slides in 2014/2015.
> 
> Having more advanced libraries definitely helps the performance but changing the up level can probably be more beneficial.
> I'm curious to know more recent advances in QE about changing internal data layout (mainly the wavefunction distribution) for better parallelization.
> Has the task group been more general rather than only FFT? Merging band group and task group if possbile...
> Though the hybrid functional or GW can be implemented with better parallel efficiency, the plain DFT part can be a severe bottle neck in the computation.
> 
> Best,
> Ye
> 
> ===================
> Ye Luo, Ph.D.
> Leadership Computing Facility
> Argonne National Laboratory
> 
> 2016-10-19 2:53 GMT-05:00 Carlo Cavazzoni <c.cavazzoni at cineca.it>:
>> Dear Sergio,
>> 
>> we are obviously addressing al that issues on different architectures with different vendors,
>> and here it come the point, architectures are not converging!
>> As you know there are two main basic designs Homogeneous and Heterogeneous (a.k.a. accelerated),
>> with some, like Intel, that oscillate between both (KNC  Heterogeneous, KNL Homogeneous, 
>> and recently announced Stratix X FPGA Heterogeneous again).
>> Is not that easy to have a code coping with all of them in an effective way, especially because
>> some of the best tools for new architectures are not standard (CUDA) and this is a real pity,
>> and make me complain with Nvidia all the time for them not supporting a standard paradigm
>> (yes I know, there is OpenACC, new OpenMP feature, OpenCL ... but CUDA remains by far more effective),
>> this is a sort of disruption for community of developers like ourself.
>> 
>> Nevertheless to reduce this complexity we recently encapsulate the two main computational kernels
>> (parallel FFT and Linear Algebra) into self contained libraries (FFTXlib and LAXlib) including a small
>> app (please read README files included in the two library) that allow one to experiment and
>> best tune all the parameters for parallelization, vectorization, tasking etc..).
>> 
>> To play with the two libraries you need to know very little about the physics of the QE,
>> and are the ideal for persons like you that need to look into optimization stuff.
>> In particular any improvements in these two libraries are immediately transferred into the QE
>> main codes (and other as well).
>> 
>> If you want to know more about our next developments, we are working with
>> non blocking MPI collectives and task based parallelism to try to overlap
>> communications and computations within the FFT.
>> Most recent (not production) advancements in FFT lib could be found at:
>> https://github.com/fabioaffinito/FFTXlib
>> 
>> 
>> Another interesting exercise could be to review the LAXlib following
>> closely the advancement in dense linear algorithms promoted by Dongarra et all
>> http://insidehpc.com/2016/10/jack-dongarra-presents-adaptive-linear-solvers-and-eigensolvers/
>> 
>> From the point of view of the paradigms we are supporting open initiatives,
>> especially in close collaboration with BSC and different standardization committees (like OpenMP),
>> or the recently announced effort promoted by AMD to open source software and drivers
>> for heterogeneous architectures: https://radeonopencompute.github.io/
>> 
>> 
>> best,
>> carlo
>> 
>> 
>> 
>> Il 18/10/2016 16:14, Sérgio Caldas ha scritto:
>>> Hi!
>>> 
>>> I'm Sérgio Caldas, an MSc student in Informatics Engineering at University of Minho, Braga, Portugal. The key area of specialisation during my master courses were on parallel computing, with a strong focus on efficient & performance engineering on heterogeneous systems. For my master thesis the theme applies these competences to computational physics, where I’m supposed to help a senior physics researcher to tune his work on the determination of electronic and optical properties of materials, using Quantum Espresso tool in our departamental cluster. This cluster has nodes with several             generations of dual multicore Xeons and some nodes with Xeon Phi (both KNC and KNL) and GPUs (both Fermi and Kepler, and soon Pascal). 
>>> 
>>> I have some queries on the QE, namely how far QE development has reached in these areas (vectorisation, data/task parallelism on both shared/distributed memory, data locality). 
>>> 
>>> For example:
>>>  - QE is already exploring vector operations (AVX/AVX-2 or AVX-512)?
>>>  - the tool is ready for multicore / many-core devices?
>>>  - how is the scheduling between multicore-devices and the accelerator  devices, such that both type of devices are simultaneously used?
>>>  - for distributed memory, the tool is already taking advantage of low-latency interconnection topologies, such as Myrinet or Infiniband?
>>>  - how can I have access to beta versions where this advanced capabilities are being explored?
>>>  - do you have suggestions of areas that still need to be improved, so that I can address those areas and improve both the quality of my work and the overall QE performance?
>>> 
>>> I would also be grateful if you could suggest documentation (preferably             papers) to get some of these answers or any other documentation to complement my knowledge on QE.
>>> 
>>> Thanking you in advance, yours sincerely
>>> Sergio Caldas
>>> 
>>> 
>>> _______________________________________________
>>> Q-e-developers mailing list
>>> Q-e-developers at qe-forge.org
>>> http://qe-forge.org/mailman/listinfo/q-e-developers
>> 
>> 
>> -- 
>> Ph.D. Carlo Cavazzoni
>> SuperComputing Applications and Innovation Department
>> CINECA - Via Magnanelli 6/3, 40033 Casalecchio di Reno (Bologna)
>> Tel: +39 051 6171411  Fax: +39 051 6132198
>> www.cineca.it
>> 
>> _______________________________________________
>> Q-e-developers mailing list
>> Q-e-developers at qe-forge.org
>> http://qe-forge.org/mailman/listinfo/q-e-developers
> 
> _______________________________________________
> Q-e-developers mailing list
> Q-e-developers at qe-forge.org
> http://qe-forge.org/mailman/listinfo/q-e-developers
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.quantum-espresso.org/pipermail/developers/attachments/20161020/3a71e9ba/attachment.html>