[gmx-users] load problem

[Stephen Chan]

Hi Justin,

My system has ~55,000 atoms. I tried reducing the number of nodes but 
the situation didn't improve. I wonder if it's related to plumed 
performance.

Stephen


On 07/22/16 22:17, Justin Lemkul wrote:
>
>
> On 7/22/16 3:58 PM, Stephen Chan wrote:
>> Hello all,
>>
>> I'm running an NPT MD on 4 nodes and each contains 28 cores. I notice 
>> the
>> performance is quite low:
>> vol 0.20! imb F  4% pme/F 0.17 step 7800, remaining wall clock 
>> time:   191 s
>>
>> I tried a couple of numbers for -npme (12, 28, 32, 40). None of these 
>> improves
>> the situation.
>>
>> I also added 'optimize-fft = yes' in the mdp file and rerun without 
>> -npme
>> option. The problem persists.
>>
>> I wonder if anyone could offer some help.
>>
>
> How many atoms are in your system?  The volume reported in the output 
> suggests that each domain is extremely small, so you may have simply 
> hit a parallelization bottleneck and you may be able to run faster 
> using fewer nodes/cores.  Is the interconnect between the nodes 
> suitably fast (e.g. InfiniBand)?
>
> -Justin
>