On 01/11/11 14:46, Christopher J.Walker wrote:
> On 01/11/11 14:43, Ewan MacMahon wrote:
>>> -----Original Message-----
>>> From: Emyr James [mailto:[log in to unmask]]
>>> Sent: 01 November 2011 14:39
>>>
>>> On 01/11/11 14:30, Ewan MacMahon wrote:
>>>> The common approach is to have per-node home directories on each
>>>> worker node's local disk, so they're effectively scratch space. There
>>>> are other approaches available depending on what storage you have
>>>> where.
>>> We have about 400GB local scratch space on each node. Will this be enough
>>> ?
>> Strictly speaking I believe the ATLAS specification is to 50Gb
>> per job slot, but they rarely seem to actually use anywhere near
>> that much. So probably yes. How many job slots are in each node?
> That requirement is based on staging data to the worker node. As Sussex
> is using StoRM, they won't need to do that, so the requirement will be
> much less. Exactly how much less, I've never managed to establish.
>
> QMUL has 500Gig drives on our dual x5650 machines - so a similar amount
> of swap space and we haven't had problems due to this, so I'd advise
> local scratch.
>
> Chris
>
Ok that sounds fine then.
|