Hi Trevor,

I would recommend starting out with NFS-- it's significantly simpler to 
configure and generally more stable than Lustre. The downside is that NFS 
doesn't do well for large sequential parallel I/O but for a small cluster 
that's probably not an issue. At the risk of starting a filesystem holy war, if 
such a day comes that NFS no-longer meets your needs I personally recommend 
looking at GPFS if you have the budget for licenses. 

Best,
Aaron

Sent from my iPhone

> On May 7, 2015, at 8:26 PM, Trevor Gale <[email protected]> wrote:
> 
> 
> Thank you for your detailed response. I think my main issue is that I’m very 
> new to Slurm, and clusters in general. I plan on setting up a global file 
> system across my desktops, and was wondering what software you would 
> recommend. I saw that the Slurm documentation mentions Lustre and NFS but was 
> just curious because I have no experience with either.
> 
> Thanks,
> Trevor 
> 
>> On May 7, 2015, at 7:28 PM, Uwe Sauter <[email protected]> wrote:
>> 
>> 
>> Trevor,
>> 
>> I don't know what your intent is or the machine you are preparing yourself 
>> for but in general login nodes and compute nodes share
>> a common filesystem, making the need to move data around (inside of the 
>> cluster) unnecessary.
>> 
>> If you really need to move data from node local space back to the login 
>> node, there are several possibilities to do so:
>> 
>> * Export some part of the login node's filesystem to your compute node.
>> * Put a SCP/RSYNC into your job script. (Make sure you're SSH keys are 
>> placed in the authorized_keys file)
>> * Run a "data mover" job that depends on your compute job (and the node 
>> where the compute job ran).
>> 
>> Likely there are more solutions to your problem. But before you go any 
>> further it'd be good if you put some thought into your
>> setup. Does it represent what you are trying to achieve?
>> 
>> Regards,
>> 
>>    Uwe
>> 
>> 
>> 
>>> Am 07.05.2015 um 15:19 schrieb Trevor Gale:
>>> 
>>> Hello,
>>> 
>>> I’m currently running one desktop computer as a controller and one as a 
>>> compute node for testing. I’m running a simple test script using salloc and 
>>> then passing the script over sbcast to my node where i execute it calling 
>>> srun. The problem I’m having is that I want the output of the programs I’m 
>>> going to run to come back to the head node (or dump there in the first 
>>> place) after execution, but all out my outputs are dropping onto the node 
>>> that they execute on. Does slurm support any method of output collection? 
>>> or is there some configuration I can change to move all the outputs to the 
>>> head node? This seems like an issue that other users would encounter, does 
>>> anyone have a good method for fixing this?
>>> 
>>> Thanks,
>>> Trevor
>>> 

Reply via email to