[gmx-users] which item is gpu wait for cpu?

[Mark Abraham]

On Thu, Sep 25, 2014 at 1:33 PM, Theodore Si <sjyzhxw at gmail.com> wrote:

> That's a breakdown of the timings that CUDA cannot always measure.
>
> What does that mean?
>

Like Szilard said, if there's more than one domain (per GPU? I forget) then
the workload each offloads to the GPU is distinct, but CUDA doesn't give us
a reliable way to assign the elapsed time separately to each domain, and
we'd need that in order to time how long the CPU cores wait for the GPU on
a per-domain basis. NB domain = rank.

We run the same test case on the same machine. This "GPU timings" part
> didn't show up before, but now it appears.
>

Ergo, you were running a different number of ranks.

Mark


> 在 9/25/2014 6:11 PM, Mark Abraham 写道:
>
>  On Thu, Sep 25, 2014 at 11:57 AM, xiexiaobin at sjtu.edu.cn <
>> xiexiaobin at sjtu.edu.cn> wrote:
>>
>>  Thanks.And I have another question want to ask.
>>> I used the command "mdrun_mpi -v -deffnm test" to run the .tpr file and
>>> find some new information in my log file:
>>>
>>>   GPU timings
>>>
>>> ------------------------------------------------------------
>>> -----------------
>>> Computing: Count Wall t (s) ms/step %
>>>
>>> ------------------------------------------------------------
>>> -----------------
>>> Pair list H2D 501 0.364 0.726 0.4
>>> X / q H2D 20001 4.153 0.208 4.5
>>> Nonbonded F kernel 18800 77.657 4.131 84.5
>>> Nonbonded F+ene k. 700 4.060 5.800 4.4
>>> Nonbonded F+prune k. 400 2.121 5.303 2.3
>>> Nonbonded F+ene+prune k. 101 0.717 7.103 0.8
>>> F D2H 20001 2.873 0.144 3.1
>>>
>>> ------------------------------------------------------------
>>> -----------------
>>> Total 91.946 4.597 100.0
>>>
>>> ------------------------------------------------------------
>>> -----------------
>>>
>>> Force evaluation time GPU/CPU: 4.597 ms/5.245 ms = 0.876.
>>>
>>> But I used to use the some command and didn't appear such
>>> information.Does
>>> anyone know why?
>>>
>>>  That's a breakdown of the timings that CUDA cannot always measure.
>>
>>
>>  I also find that the time consuming of Nonbonded F kernel is 77.657s but
>>> the time consuming of force is only 55.263s.Does the force item include
>>> non-bonded force?
>>>
>>
>> No, it's non-PME work being done on the CPU (e.g. bonds, restraints,
>> free-energy short-ranged)
>>
>> Mark
>>
>> If it was true,why the Nonbonded F kernel takes more time than the force ?
>>
>>>
>>>
>>> xiexiaobin at sjtu.edu.cn
>>>
>>> From: Szilárd Páll
>>> Date: 2014-09-24 17:45
>>> To: Discussion list for GROMACS users
>>> CC: gromacs.org_gmx-users
>>> Subject: Re: [gmx-users] which item is gpu wait for cpu?
>>> On Wed, Sep 24, 2014 at 4:57 AM, xiexiaobin at sjtu.edu.cn
>>> <xiexiaobin at sjtu.edu.cn> wrote:
>>>
>>>> I think that the items (wait GPU nonlocal) and (wait GPU local) in log
>>>>
>>> file are the time consuming that cpu wait for gpu. And Does anyone know
>>> that which item is the time consuming that gpu wait for cpu in log file?
>>>
>>> Due to a CUDA runtime limitation, that can not measured except for the
>>> no domain-decomposition case (where it is expressed as overlap).
>>>
>>>
>>>> xiexiaobin at sjtu.edu.cn
>>>> --
>>>> Gromacs Users mailing list
>>>>
>>>> * Please search the archive at
>>>>
>>> http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_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-users or
>>>>
>>> send a mail to gmx-users-request at gromacs.org.
>>> --
>>> Gromacs Users mailing list
>>>
>>> * Please search the archive at
>>> http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_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-users or
>>> send a mail to gmx-users-request at gromacs.org.
>>> --
>>> Gromacs Users mailing list
>>>
>>> * Please search the archive at
>>> http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_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-users or
>>> send a mail to gmx-users-request at gromacs.org.
>>>
>>>
> --
> Gromacs Users mailing list
>
> * Please search the archive at http://www.gromacs.org/
> Support/Mailing_Lists/GMX-Users_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-users or
> send a mail to gmx-users-request at gromacs.org.
>