[GE users] h_vmem? virtual_free? mem_free?

Sebastian Stark stark at tuebingen.mpg.de
Wed Apr 13 16:18:54 BST 2005


I must say I'm totally confused the more I read the documentation, howtos and 
mailinglist archives.

What I'm trying to do is simple: avoid memory oversubscription, especially in 
the case where a users submits a lot of jobs which all need roughly the 
amount memory a single host can offer.

I've read about simply using the mem_free parameter as a hard resource 
request. Does not work because if a job is dispatched sge does not subtract 
the memory it needs (-hard -l mf=8G) immediately. So another job is 
dispatched right after the first one (until all slots are occupied) and then 
the swapping begins. The Job needs some time to fill the 8G, I don't know how 
long in advance. This behaviour does not change when I use "qconf -mc" and 
set the requestable flag to YES for the mem_free resource.

Then I read about using virtual_free in a way which sounded completely absurd 
to me: First I should make virtual_free requestable and than _define_ it to 
be the same as the available _real_ memory in the complex_values list in each 
exec hosts configuration. When I do this it is hardly "virtual" anymore. And 
I do not understand what's the difference to mem_free then.

In the pdf manual I also read about the h_vmem parameter which can be used to 
accomplish that. What's the difference to virtual_free?

I'm using n1ge 6u3 and I have a single cluster queue.


Thanks for any help,
Sebastian

-- 
Sebastian Stark -- http://www.kyb.tuebingen.mpg.de/~stark
Max Planck Institute for Biological Cybernetics
Spemannstr. 38, 72076 Tuebingen
Phone: +49 7071 601 555 -- Fax: +49 7071 601 552

---------------------------------------------------------------------
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