Am 15.01.2012 um 08:56 schrieb mahbube rustaee:

> Hi all,
> 
> A case that a requested resource  issue control of other resource is request 
> h_vmem.
> requested h_vmem cause some core of node be locked! 
> I have two nodes : Xeon , AMD
> Xeon nodes are 24 slots and 23GRAM. (about 1 GIG per core)
> AMD nodes are 48 slots and 90G RAM. (about 2 GIG per core)
> 
> I 'd like users could request memory as much as share of  requested cores 
> (e.g. 12 core could request max 12G on Xeon or max 24G on AMD) . as a result, 
> cores cann't be locked caused by requested memory.

Before you described always your observation. If you want to achieve just this 
limit: forget all what is set up and define in the queue:

$ qconf -sq all.q
...
h_vmem                INFINITY,[@xeon=1G],[@amd=2G]

This limit will be applied per slot. It's even not necessary for the users to 
request anything now, as the limit will always be enforced and requesting a 
lower limit won't grant access to any other user to use it, as they can't 
request more than this set limit.

-- Reuti


> Pardon me, I asked  such question into different form but I didn't got 
> suitable answer.suggestion was:
> 1- limit h_vmem on hosts.
>  In that case, h_vmem is per host whereas requested h_vmem via job is per 
> slot.
> 
> 2- set appropriate value (memory per core) on consumable resource .
> It cann't be done because  memory per slot (h_vmem) is different on Xeons and 
> AMDs).
> 
> 2- control requested via JSV.
>  this cann't be operational because jsv could test fixed number on requested 
> h_vmem whereas max amount of memory per host  is determined after running and 
> via scheduling and this amount is different on AMD nodes and Xeon nodes.
> 
> Any suggestion would be appreciated.
> _______________________________________________
> users mailing list
> [email protected]
> https://gridengine.org/mailman/listinfo/users


_______________________________________________
users mailing list
[email protected]
https://gridengine.org/mailman/listinfo/users

Reply via email to