[gmx-developers] combine neighborlist

[Mark Abraham]

Hi,

Also, the setup is intended to allocate and first-touch memory on each
OpenMP thread so that (particularly) the output buffers do not suffer from
false sharing issues. This would be a problem with a naive approach to
"make one list and split it over all cores."

Mark

On Wed, Jul 8, 2015 at 1:08 AM Szilárd Páll <pall.szilard at gmail.com> wrote:

> On Wed, Jul 8, 2015 at 12:39 AM, Sikandar Mashayak <mashayak at votca.org>
> wrote:
>
>> Hi,
>>
>> I am trying to understand how a SIMD-based kernel distributes the work
>> across OpenMP threads. I found that, given n threads, n neighborlists are
>> built, each OpenMP thread is assigned a neighborlist and for each
>> neighborlist outer and inner loop computations are performed using SIMD
>> instructions.
>>
>> My question is, whether building n neibhorlists and assigning 1 list to
>> each thread is same (from performance considerations) as creating 1
>> neighborlist and dividing i-clusters among n threads (nci/n per thread)?
>>
>
> Not exactly. Lists are of different length and when N lists are produced
> for N non-bonded compute tasks, the search tries to generated balanced
> work. Splitting using the simple static nci/N formula will give imbalanced
> chunks of work.
>
> You never said what are you trying to accomplish, how are you going to use
> the single 4x4 list? You may be able to avoid some pitfalls by sharing the
> bigger picture.
>
> --
> Szilárd
>
>
>> Thanks,
>> Sikandar
>>
>> On Tue, Jun 30, 2015 at 1:57 PM, Sikandar Mashayak <mashayak at votca.org>
>> wrote:
>>
>>> Thanks Szilard,
>>>
>>> Yeah, I guess with some coding merging neighborlists into a single
>>> simple CPU-style list is possible. For now, I am manually setting number of
>>> neighborlist to 1 independent of openmp threads.
>>>
>>> Thanks,
>>> Sikandar
>>>
>>> On Wed, Jun 24, 2015 at 8:00 PM, Szilárd Páll <pall.szilard at gmail.com>
>>> wrote:
>>>
>>>> Hi,
>>>>
>>>> I think it should be supported if not out of the box, perhaps with
>>>> small code changes. I know we have discussed decoupling the search
>>>> form the force compute to allow executing them on different number of
>>>> executors/tasks. Also, I'm not entirely sure if the implementation
>>>> allows searchig on N tasks and merging simple CPU-style lists into a
>>>> single list output.
>>>>
>>>> Try looking into the nbnxn_init_pairlist_set() function which gets
>>>> passed a boolean that tells whether to combine lists or not.
>>>>
>>>> Note that you can control the number of tasks (=OpenMP threads) in
>>>> different parts of the code through the gmx_omp module which reads
>>>> environment variables in the form of GMX_*_NUM_THREADS.
>>>>
>>>> Cheers,
>>>> --
>>>> Szilárd
>>>>
>>>>
>>>> On Thu, Jun 25, 2015 at 1:06 AM, Sikandar Mashayak <mashayak at votca.org>
>>>> wrote:
>>>> > Hi,
>>>> >
>>>> > I want to use pairlist with M = 4 and N = 4 i and j cluster sizes. I
>>>> > understand, that with OpenMP threads enabled, gromacs builds ntomp,
>>>> i.e.,
>>>> > nthreads, neighborlists. I understand, for 8*8*8 kernel type, these
>>>> > neighborlists are combined and superclusters are built.
>>>> >
>>>> > I am wondering, is there a way to combine all the nthreads
>>>> neighborlist into
>>>> > one for simple 4*4 type neighborlist?
>>>> >
>>>> > Thanks,
>>>> > Sikandar
>>>> >
>>>> > --
>>>> > Gromacs Developers mailing list
>>>> >
>>>> > * Please search the archive at
>>>> > http://www.gromacs.org/Support/Mailing_Lists/GMX-developers_List
>>>> before
>>>> > posting!
>>>> >
>>>> > * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>>>> >
>>>> > * For (un)subscribe requests visit
>>>> >
>>>> https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-developers
>>>> or
>>>> > send a mail to gmx-developers-request at gromacs.org.
>>>> --
>>>> Gromacs Developers mailing list
>>>>
>>>> * Please search the archive at
>>>> http://www.gromacs.org/Support/Mailing_Lists/GMX-developers_List
>>>> before posting!
>>>>
>>>> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>>>>
>>>> * For (un)subscribe requests visit
>>>> https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-developers
>>>> or send a mail to gmx-developers-request at gromacs.org.
>>>
>>>
>>>
>>
>> --
>> Gromacs Developers mailing list
>>
>> * Please search the archive at
>> http://www.gromacs.org/Support/Mailing_Lists/GMX-developers_List before
>> posting!
>>
>> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>>
>> * For (un)subscribe requests visit
>> https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-developers
>> or send a mail to gmx-developers-request at gromacs.org.
>>
> --
> Gromacs Developers mailing list
>
> * Please search the archive at
> http://www.gromacs.org/Support/Mailing_Lists/GMX-developers_List before
> posting!
>
> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>
> * For (un)subscribe requests visit
> https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-developers
> or send a mail to gmx-developers-request at gromacs.org.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://maillist.sys.kth.se/pipermail/gromacs.org_gmx-developers/attachments/20150708/877aac76/attachment.html>