From: Josh Vermaas (vermaas2_at_illinois.edu)
Date: Thu Apr 28 2016 - 10:09:28 CDT
Hi Sourav,
It is just not how NAMD works. ACEMD or AMBER can both run as you
describe, in that there is no communication between GPU and CPU once the
job has started. NAMD wasn't designed this way, and requires a CPU
associated with the process to deal with bonded interactions.
-Josh
On 04/28/2016 10:04 AM, Sourav Ray wrote:
> Hello Brian
>
> Is there any alternative to CUDA version that might support a GPU
> only simulation? I checked the mailing list and couldn't find a
> relevant answer to this.
>
> Thanks and regards
> Sourav
>
> On Thu, Apr 28, 2016 at 8:30 PM, Brian Radak <bradak_at_anl.gov
> <mailto:bradak_at_anl.gov>> wrote:
>
> No. That's not how the NAMD-CUDA implementation is designed. Many
> operations in normal MD are only ever performed on the CPUs.
>
> Brian
>
>
> On 04/28/2016 09:40 AM, Sourav Ray wrote:
>> Hello
>>
>> Is there any way I can use the NAMD-CUDA version to run
>> simulations without using any of the CPU cores and restricting
>> the simulation to only GPU unit? I tried the following command:
>>
>> NAMD_2.11_Linux-x86_64-multicore-CUDA/namd2 +p0 +devices 0
>> test.conf > test.log
>>
>> It doesn't work and putting +p0 shows the following error:
>>
>> pthread_create: Resource temporarily unavailable
>> Segmentation fault (core dumped)
>>
>> Please let me know if there is a workaround.
>>
>> Thanks and regards
>> Sourav
>
> --
> Brian Radak
> Postdoctoral Appointee
> Leadership Computing Facility
> Argonne National Laboratory
>
> 9700 South Cass Avenue, Bldg. 240
> Argonne, IL 60439-4854
> (630) 252-8643
> brian.radak_at_anl.gov <mailto:brian.radak_at_anl.gov>
>
>
This archive was generated by hypermail 2.1.6 : Sun Dec 31 2017 - 23:20:22 CST