Hi,
> 1) I did a quick benchmark on a system with 16 3.2 GHz core-I7's and 4
> gtx1080's using 1000 super-res K2 movies.
> 4 mpi's with 1 gpu each is faster (2.6 hours) than 1 mpi with 4 gpus
> (3.5 hours) on this system.
Thank you very much for testing.
Although it might be slower, I updated the wrapper (commit 5dc615767) to
allow
use of more than one GPU per MPI process by explicitly specifying GPUs.
For example, specify "0,1:2,3" for 2 MPI processes, 2 GPUs each.
> I'd mistakenly assumed that if you simply
> specified 4 mpi's and let Relion automatically choose gpu's,
This should work, as long as you are running on a single machine.
If more than two machines are present, it is safer to assign
explicitly. Suppose you want to run 2 MPI processes times
2 nodes and each node has 2 GPUs. If MPI ranks are assigned
as (0, 1), (2, 3), they will use GPU (0, 1), (0, 1), which is good.
However, if MPI ranks are assigned as (0, 2), (1, 3), they will use
GPU (0, 0), (1, 1). In this case, specify "--gpu 0:0:1:1".
Best regards,
Takanori Nakane
> 1) I did a quick benchmark on a system with 16 3.2 GHz core-I7's and 4
> gtx1080's using 1000 super-res K2 movies.
> 4 mpi's with 1 gpu each is faster (2.6 hours) than 1 mpi with 4 gpus
> (3.5 hours) on this system.
>
> 2) Ialso? found the reason I originally concluded this didn't work. As
> Alpay pointed out, if running from the gui you must manually assign
> gpu's using ":" notation. I'd mistakenly assumed that if you simply
> specified 4 mpi's and let Relion automatically choose gpu's, it would
> default to sequential assignment as it does with 2D and 3D
> classification. That doesn't seem to work in this case.
>
> Regards,
> -jh-
>
> On 12/15/18 8:57 PM, Daniel Asarnow wrote:
>>
>> Is assigning all 4 GPUs to one MotionCor2 faster than
>> running 4 instances of MotionCor2?
>>
>>
>> It's supposed to be, and that's how we use it at UCSF for on-the-fly
>> correction. I'll double check our benchmarks.
>>
>> Best,
>> -da
>>
>> ------------------------------------------------------------------------
>>
>> To unsubscribe from the CCPEM list, click the following link:
>> https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=CCPEM&A=1
>>
>
> --
> John M. Heumann
> Department of Molecular, Cellular, and Developmental Biology
> 347 UCB, University of Colorado
> Boulder, CO 80309-0347
>
> ########################################################################
>
> To unsubscribe from the CCPEM list, click the following link:
> https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=CCPEM&A=1
>
########################################################################
To unsubscribe from the CCPEM list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=CCPEM&A=1
|