Hi,
Burke, S (Stephen) wrote:
>LHC Computer Grid - Rollout
>
>
>>[mailto:[log in to unmask]] On Behalf Of Fabien Wernli
>>
>>
>said:
>
>
>>>Do you have a batch system that can cope with that?
>>>
>>>
>>If it doesn't we can make it cope with it
>>
>>
>
>Maybe you can, but most people are using PBS and I'd guess it would be
>pretty hard to get that to cope.
>
>
>
>>>And you'd also need to avoid NFS mounts
>>>[...]
>>>
>>>
>>do you have per user/DN NFS mounts?
>>
>>
>
>My point is that if you have any writeable space mounted across all the
>WNs (especially home directories) you would get interference between
>jobs for different users running under the same account.
>
>
>
>>>[...]
>>>or implicit ssh logins.
>>>
>>>
>>?
>>
>>
>
>At the moment the system usually uses scp to push files back from the WN
>to the CE, and may allow ssh between WNs for MPI jobs, none of which
>would work if the same accounts on different nodes belong to different
>users.
>
>
>
I thought to have noticed that some systems didn't used scp and ssh to
do such things. Like executing a plain cp from WN to CE which doesn't
require ssh-keys.
One can think of the method of working with MPI in such a way that one
job will be mapped to one poolaccount across X requested MPI-enabled
nodes (where X is requested from the Job specification).
>Stephen
>
>
Oscar
|