[GE users] Is it possible to set h_vmem to the same value as mem_free

Michael Lindig michael.lindig at informatik.tu-chemnitz.de
Wed Jul 2 00:22:42 BST 2008


thanks this helps,

I added per 'qconf -me nodex' h_vmem=8G to the complex_values and modified the
complex definition for h_vmem to:

h_vmem   h_vmem     MEMORY      <=    FORCED      YES        0        0

Now it works as expected ;)! It would be nice if I could set the complex_values
for h_vmem per RQS like for slots:

limit        hosts {*} to slots=$num_proc

what I want:

limit        hosts {*} to h_vmem=$num_proc*2G

... but I read that dynamic RQS only possible for INT and DOUBLES :(!

Thanks again for help,
Michael
> Hi,
> 
> Am 01.07.2008 um 23:33 schrieb Michael Lindig:
> 
> > I have a "small" cluster where each node has only 1GByte of swap
> > and 8GB of ram
> > (each has 4 CPUs) only 2 nodes have 32 GB ram and 1GB swap (and 16
> > CPUs). The
> > h_vmem value is set to required, but it seams that the gridengine
> > (I use 6.1u4)
> > does not check if the node has the capability to fulfil the
> > resources. If I
> > request h_vmem=16G then I will scheduled also to nodes with 8GB ram!?
> 
> yes, it's just a limit (something you could also set with "ulimit" in
> the shell). It will even set the limit to such a "nonsense" value on
> certain nodes.
> 
> >
> > Is it possible to request mem_free=2G and set automatically the
> > h_vmem=mem_free?
> >
> > Or is there another possibility to prevent the nodes for swap usage?
> >
> > Our problem is that any user can allocate more memory then
> > available and bring
> > the node to unmanage state if it starts swapping:(!
> 
> To share it across all jobs on a node:
> 
> http://gridengine.sunsource.net/servlets/ReadMsg?
> listName=users&msgNo=24174
> 
> http://gridengine.sunsource.net/servlets/ReadMsg?
> listName=users&msgNo=12027
> 
> 
> If you just want just a limit, you could even set the maximum which
> can be requested in the queue definition:
> 
> h_vmem INFINITY,[@smallones=8g],[@bigones=32g]
> 
> But many jobs might request too much memory in total on a node this
> way. Nevertheless we use this in addition to making h_vmem
> consumable: one a queue level the limit is 12g (i.e. per job), on a
> host level it's the installed 16g (not to many jobs will be scheduled
> to one and the same node and exhaust the memory).
> 
> HTH - Reuti
> 
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: users-unsubscribe at gridengine.sunsource.net
> For additional commands, e-mail: users-help at gridengine.sunsource.net
> 



---------------------------------------------------------------------
To unsubscribe, e-mail: users-unsubscribe at gridengine.sunsource.net
For additional commands, e-mail: users-help at gridengine.sunsource.net




More information about the gridengine-users mailing list