Just wondering, how often do you run jobs that use 2 or more GPUs? As far as I know, GPUs don't share memory with the system memory directly, and it can be even harder for 2 GPUs in a system to share memory. Does CUDA or OpenCL handle the data transfer details for the programmer?
-Ron ----- Original Message ----- From: William Hay <[email protected]> To: Ron Chen <[email protected]> Cc: Rayson Ho <[email protected]>; "[email protected]" <[email protected]> Sent: Thursday, May 24, 2012 3:02 PM Subject: Re: [gridengine users] cgroups Integration in OGS/GE 2011.11 update 1 On 24 May 2012 17:27, Ron Chen <[email protected]> wrote: > I think it is a great way to handle GPU permissions, and I sent your method > of managing GPU permissions to a GPU & clustering discussion list. > > However, I assumed that you were using multiple queue instances per node, and > each queue logically owns the GPU. Was my assumption correct? > > -Ron No because that would make it trickier for a single slot job to request both GPU on a node. We use lockfiles manipulated in the prolog and epilog to keep track of which GPU are allocated to which job. William _______________________________________________ users mailing list [email protected] https://gridengine.org/mailman/listinfo/users
