[torquedev] nodes, procs, tpn and ncpus

"Mgr. Šimon Tóth" SimonT at mail.muni.cz
Fri Jun 11 14:18:26 MDT 2010


On 11.6.2010 21:45, Martin Siegert wrote:
> On Fri, Jun 11, 2010 at 01:24:15PM -0600, Ken Nielson wrote:
>> On 06/11/2010 12:19 PM, "Mgr. Šimon Tóth" wrote:
>>> On 11.6.2010 18:41, Martin Siegert wrote:
>>>    
>>>>
>>>> It all depends on whether we want to break the meaning of nodes=N being
>>>> totally equivalent to nodes=N:ppn=1.
>>>> 1) if we stick to the meaning of nodes=N just being a shorthand for
>>>>     nodes=N:ppn=1 then -l nodes=4+procs=4 makes just as much sense as
>>>>     nodes=4:ppn=2+procs=6, i.e., there is no problem.
>>>> 2) if we change the meaning of nodes=N to become "give me exclusive
>>>>     access to N nodes" then we need to define what -l nodes=N+procs=M
>>>>     means.
>>>>
>>>> I actually vote for (1). E.g., what is the default if neither nodes nor
>>>> procs is specified? Currently that is nodes=1:ppn=1, i.e., the default
>>>> for nodes is 1 and the default for ppn is one. That logical extension is
>>>> that if only nodes is specified then the default value for ppn should be
>>>> used. I.e., nodes=N means nodes=N:ppn=1.
>>>> But that leaves us with the problem of how to specify exclusive access
>>>> to a node. What about a keyword ALL as in nodes=4:ppn=ALL ? It is a
>>>> logical extension of the current syntax and is easy to understand.
>>>> And then things like -l nodes=1:ppn=ALL+procs=42 make total sense
>>>> (at least to me).
>>>>      
>>> That's not good enough. How do you tell Torque that you want at least a
>>> 4 CPU node, but when you get it, then you want all the CPUs?
>>>
>>>    
>>>    
>> How about just add #exclusive or other appropriate keyword.
> 
> That would work for me.

That's why I have all nodes configured to np=1. When a non-shared job
enters the node, it is switched in the job-exclusive state. When a
shared job enters the node, then it just eats a portion of the resources
on the node.

PBSPro supports #excl flag.

-- 
Mgr. Šimon Tóth

-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/pkcs7-signature
Size: 3366 bytes
Desc: S/MIME Cryptographic Signature
Url : http://www.supercluster.org/pipermail/torquedev/attachments/20100611/66b9ebb8/attachment.bin 


More information about the torquedev mailing list