Questions of QMCPACK

[jim8001...@gmail.com]

Hi,
I would like to ask some questions of QMCPACK.
First question is that, is there any visualization application for QMCPACK?
I couldn't find it.
Second, there are several tutorials on the website of QMCPACK, but some examples of them can not use GPU to accelerate, will them appear in this competition? If they will, should we need to determine that which can use GPU to accelerate or not ?

Thanks
Jim Chen

[Kim, Jeongnim]

On 10/18/12 6:24 AM, "jim8001...@gmail.com" <jim8001...@gmail.com>
wrote:

>Hi,
>I would like to ask some questions of QMCPACK.
>First question is that, is there any visualization application for
>QMCPACK?
>I couldn't find it.

No, there is no special vis need or application for QMCPACK. The analysis
of the data and performance will be done by the scripts in utils.

>Second, there are several tutorials on the website of QMCPACK, but some
>examples of them can not use GPU to accelerate, will them appear in this
>competition? If they will, should we need to determine that which can use
>GPU to accelerate or not ?

I will specify the target problems on the web site
http://qmcpack.cmscc.org/how-to-on-hpc/sc12-student-cluster-competition

And add more instructions.

All the examples can run on both GPUs and CPUs.


Currently, the input files for two examples are available.
* 3x3x1 graphite (small)
* 4x4x1 graphite (big)

Third one will be posted soon.


>
>Thanks
>Jim Chen
>
>--
>You received this message because you are subscribed to the Google Groups
>"SCC12" group.
>To post to this group, send email to sc...@googlegroups.com.
>To unsubscribe from this group, send email to
>scc12+un...@googlegroups.com.
>Visit this group at http://groups.google.com/group/scc12?hl=en.
>
>

[Jessica Dudoff]

Hello,

I have a few additional questions regarding QMCpack.

• QMCPACK operates in two very different computational regimes depending on the nodal approximation that is used:  an analytic form (free particle nodes, atomic orbitals) vs. a DFT-derived dataset read in from file.  To avoid having to optimize for all possible use cases, can any guidance be given as to which of these two types of systems we can expect?
  
• Will the number of steps and blocks be set at the outset, or is the expectation that teams will have to choose an appropriate number to achieve a given statistical error bar?
  

Thank you,

Jessica Dudoff

[Kim, Jeongnim]

To be able to run on both CPU and GPU clusters, the examples will use the prepared orbitals in a HDF5 file.
Each run will have

* initialization : read from xml and hdf5, FFT transform using FFTW and set up interpolation table using einspline library
* Vmc : warmup block to generate DMC samples
* Dmc : this is the main computation section

Critical routines for the optimal performance are in src/einspline directory. Optimized HDF5 and FFTW libraries can improve the initialization performance.
However, the input parameters will be chosen so that the initialization is not critical for the overall performance, I.e., the time-to-solution.

Each example will use the fixed number of MC steps, I.e., blocks and steps are fixed.

Each team can choose the total workloads (samples) and the number of parallel processing units (MPI tasks and OpenMP threads). A minimum workload will be specified for each example.

I will post the updated instructions by Monday next week.

Bests,
Jeongnim

From: Jessica Dudoff <jdu...@gmail.com<mailto:jdu...@gmail.com>>
Reply-To: "sc...@googlegroups.com<mailto:sc...@googlegroups.com>" <sc...@googlegroups.com<mailto:sc...@googlegroups.com>>
Date: Thursday, October 18, 2012 5:11 PM
To: "sc...@googlegroups.com<mailto:sc...@googlegroups.com>" <sc...@googlegroups.com<mailto:sc...@googlegroups.com>>
Cc: "jim8001...@gmail.com<mailto:jim8001...@gmail.com>" <jim8001...@gmail.com<mailto:jim8001...@gmail.com>>
Subject: SCC12 Re: Questions of QMCPACK

Hello,

I have a few additional questions regarding QMCpack.

* QMCPACK operates in two very different computational regimes depending on the nodal approximation that is used: an analytic form (free particle nodes, atomic orbitals) vs. a DFT-derived dataset read in from file. To avoid having to optimize for all possible use cases, can any guidance be given as to which of these two types of systems we can expect?

* Will the number of steps and blocks be set at the outset, or is the expectation that teams will have to choose an appropriate number to achieve a given statistical error bar?

Thank you,

Jessica Dudoff

On Thursday, October 18, 2012 3:24:15 AM UTC-7, jim8001...@gmail.com<mailto:jim8001...@gmail.com> wrote:
Hi,
I would like to ask some questions of QMCPACK.
First question is that, is there any visualization application for QMCPACK?
I couldn't find it.
Second, there are several tutorials on the website of QMCPACK, but some examples of them can not use GPU to accelerate, will them appear in this competition? If they will, should we need to determine that which can use GPU to accelerate or not ?

Thanks
Jim Chen

--
You received this message because you are subscribed to the Google Groups "SCC12" group.

To post to this group, send email to sc...@googlegroups.com<mailto:sc...@googlegroups.com>.
To unsubscribe from this group, send email to scc12+un...@googlegroups.com<mailto:scc12+un...@googlegroups.com>.

[jim8001...@gmail.com]

jnkim於 2012年10月20日星期六UTC+8上午5時27分59秒寫道：

Hi,

Thanks, but there is a parameter I want to ask, it's in dmc section "targetwalker", will it fixed?

Thanks,
Jim Chen

[Kim, Jeongnim]

I updated the wiki on the problem at
http://qmcpack.cmscc.org/how-to-on-hpc/sc12-student-cluster-competition
All the inputs and outputs are available via subversion

The minimum "samples" in the vmc section is 8000. Do not set targetwalkers
in DMC, although setting it again would not change the run.

Jeongnim


On 10/23/12 10:16 AM, "jim8001...@gmail.com"

>>scc12+un...@googlegroups.com<mailto:scc12+unsubscribe@googlegroups.

>>com>.
>>
>> Visit this group at http://groups.google.com/group/scc12?hl=en.
>
>Hi,
>
>Thanks, but there is a parameter I want to ask, it's in dmc section
>"targetwalker", will it fixed?
>
>Thanks,
>Jim Chen
>

>--
>You received this message because you are subscribed to the Google Groups
>"SCC12" group.

>To post to this group, send email to sc...@googlegroups.com.

>To unsubscribe from this group, send email to

>scc12+un...@googlegroups.com.

[chenzh...@gmail.com]

Hi I have a question, that will each team have the same "samples" despite we have different numbers of CPU cores?
In the website "samples" is supposed to be "MPI TASKS"*"OPENMP THREADS"*"VMCBLOCKS".
Thanks
Nick

在 2012年10月18日星期四UTC+8下午6时24分15秒，jim8001...@gmail.com写道：

[Kim, Jeongnim]

For GPUs, the number of CPU cores does not matter, OMP_NUM_THREADS=1. It
is best to set
VMCWALKERS >= 32

The larger this number is (128 or 256), the better performance one will
get.

Jeongnim

On 10/23/12 9:32 PM, "chenzh...@gmail.com" <chenzh...@gmail.com>
wrote:

[jim8001...@gmail.com]

hi,

There is another question,
when the application was running, it occured this error.

Fatal Error. Aborting at WalkerControlBase::sortWalkers
[cli_1]: aborting job:
application called MPI_Abort(MPI_COMM_WORLD, 1) - process 1

Is there any solution of tis fatal error?

Thanks,
Jim Chen

[Kim, Jeongnim]

It is most likely that samples per core or GPU is too low.
Send me the main input file and the run parameters.
Jeongnim

On 10/26/12 11:31 PM, "jim8001...@gmail.com"

[Kim, Jeongnim]

Hi,

Upon requests from the participants, a snapshot is available at
http://qmcpack.googlecode.com/files/qmcpack-20121029.tgz

It is an experimental version with major changes in the CUDA
implementations. Very limited tests have been carried out.

Use it at your own risk.

Cheers,
Jeongnim

[jiazho...@gmail.com]

在 2012年10月30日星期二UTC+8上午2时02分31秒，jnkim写道：

Hi

I've tried to do some work on this snapshot.

It seems to be used on KEPLER GPUs although FERMI GPUs can also do. But I can't finish compiling it even with CUDA5(Syntax error). I just can't figure out it. Will it be very valuable to still work on it?

Besides, is there any considerable differences between it and the previous version on CPU codes？

Thanks

Nick

[Kim, Jeongnim]

Nick,

No, there is no need to use this version, especially if you have
encountered CUDA errors. Someone asked for a developer's version and I
wanted to be fair for all the participants.
The changes in CPU version is very minimal.

Jeongnim


On 10/30/12 4:56 AM, "jiazho...@gmail.com" <jiazho...@gmail.com>
wrote: