On Mar 12, 2013, at 7:41 AM, Serge A. Salamanka <salsa-...@tut.by> wrote:
> On 06.03.2013 15:15, Sebastien Goasguen wrote: >> If you don't need to multiplex regular hpc jobs and VMs then things gets >> easier. Just build a CloudStack cloud, and setup your HPC cluster via >> templates. One template for a head node and one template for the compute >> nodes. Put MPI or whatever libraries you need and your are up. > > I don't need to multiplex HPC jobs and VMs. > This is exactly the setup I want to make. > This way we can provide users the ability to create their own HPC > clusters with their own templates for worker nodes as configurable > environment. > > >> CloudStack won't help with the actual parallelizing, that's not what > it's made for. > > CloudStack can help to create highly configurable virtual infrastructure > for HPC jobs that can be used by organizations that care about strong > privacy of the submitted information and by those individuals that want > root access to their HPC nodes. > >> If you wanted an optimized VM instance, for true HPC, you would need > to set it up on hypervisors with PCI passthrough, some type of low > latency network like infiniband. In addition you could have GPU > instances (like on Amazon), if you had GPUs on your hardware. > > Infiniband and GPU nodes are available. The task is to make them > available through CloudStack. > > I would appreciate more information on configuring optimized VMs. > The way I would do it is to first work without CloudStack, focusing on the hypervisor layer. You need to be able to use infiniband properly through a VM, I believe there are special drivers, but could be that virtio is now enough to get the low latency. Once you know how to do that, you can create a service offering in cloudstack. For GPUs, I still believe you need to setup a pass through (quick search): http://superuser.com/questions/427918/virtual-machine-with-dedicated-gpu-is-it-possible-private-gpu-cloud and that you will have to dedicate GPUs to a single user. With bare metal functionality in CS 4.1, you might want to look into provisioning a bare metal cluster from the nodes that have GPUs. > #Serge
