Agree with Rémi but adding a few comments below.

On Sat, 23 May 2015 08:49:13 -0700
Rémi Palancher <[email protected]> wrote:

> 
> Le 23/05/2015 01:29, Trevor Gale a écrit :
> >
> > Hello all,
> >
> > I am new to slurm and HPC in general, and am trying to learn how
>  > other pieces of software interact with slurm. I was wondering if
>  > any changes need to be made to the slurm.conf when using an
>  > infiniband network (e.g. using LIDs or GUIDs instead of IPs for
>  > node addresses) and if you use the MLNX_OFED inifiniband software
>  > available on mellanox’s website. Also I was wondering what MPI
>  > works best with slurm and an IB network.
> 
> Hi Trevor,
> 
> Actually, Slurm does not do much with IB. Whatever the underlying 
> network (IB, Ethernet, and so on), Slurm components require IPv4 
> addresses to communicate, it is not able to use low-level protocols 
> under IP-o-IB.

That is all true but slurm does handle topology for the IB network. You
(can) make a topology config file and this allows better-than-random job
placements (YMMV).

> If you have Mellanox hardware it is better to use Mellanox flavor of 
> OFED to get all the performances with last Mellanox technologies
> (MXM, FCA, etc) and get technical support from them.

Customers may value the simplicity of updates when using the dist.
provided ib stack over the latest and greatest features of mlnxofed.
 
> Most major implementations of MPI (Open MPI, MPICH, Intel MPI and 
> probably others) are tightly integrated with Slurm, esp. when PMI 
> support is enable.

There are many levels of integration with slurm including but not
limited to:

* mpi does not understand slurm but is fed an auto-created hostfile
* mpi understands slurm environment and launches ranks on allocated
  nodes but via, maybe, ssh (or other non-slurm mech.)
* mpi understands slurm environment and uses srun to launch on other
  nodes (like for example intelmpi "mpiexec.hydra -bootstrap slurm").
* srun used to run mpi binary and remote process launch via PMI

/Peter K

Reply via email to