Multiple K80 GPU's

[Richard Casey]

Hi,

We're running NVBIO benchmarks on a cluster with K80 GPU cards.  The cluster has multiple compute nodes, with a single K80 card per node.  The single GPU results look very good.  We would like to run NVBIO on multiple GPU's.  In SLURM, for 2 K80 GPU's (1 K80 per node), we're trying

#SBATCH --nodes=2 --ntasks-per-node=1

#SBATCH --gres=gpu:2

#SBATCH --constraint=K80

Does NVBIO automatically detect the presence of multiple GPU's on multiple nodes?  Does it automatically run a job across multiple GPU's? Or do we have to set other parameters to do this?

thx.  

[Jacopo Pantaleoni]

Hi,

You have ti specify each gpu separatelyon the command line, e.g. nvBowtie --device 0 --device 1 ...

--
You received this message because you are subscribed to the Google Groups "nvbio-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to nvbio-users...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Richard Casey]

Hi,

We're now running nvBowtie on another multi-GPU system.  It has several compute nodes, each node has a single K20 card, and nodes are connected with Infiniband FDR.  On each compute node the device ID=0.  If we try multi-GPU runs (say, 2 GPU's) with:

nvBowtie --device 0 --device 1

it returns the error message

FMIndexDataDevice: invalid device ordinal

caught a nvbio::cuda_error exception:

invalid device ordinal

It probably doesn't like device ID=1.  Are we running this correctly across multiple compute nodes?