[GE users] Differences between NODES and SLOTS

Bevan C. Bennett bevan at fulcrummicro.com
Mon Mar 5 22:58:32 GMT 2007


    [ The following text is in the "ISO-8859-1" character set. ]
    [ Your display is set for the "ISO-8859-10" character set.  ]
    [ Some special characters may be displayed incorrectly. ]

Daniel Ruiz wrote:
> Hi!!
> 
> We are going to install a new cluster with this configuration:
>   ---> 1 management node with 1 CPU Xeon Dual-Core
>   ---> 16 compute node with 2 CPUs Xeon Dual-Core
> 
> We have the following questions:
> 
> 1. If we have 16 compute nodes with 2 CPUs Dual-Core (total=64 cores),
> by default, if we run "qstat -f", we should see 16 queuenames. But, if
> each node has 2 CPUs and each CPU has 2 cores, how can we get 64
> queuenames? Is it possible?

What you should see is one queue with 16 "queue instances", one per physical
system. You then manually set the correct number of slots in each queue instance
to the correct number of cores.

For example, we have several "dual dual core" systems (like you have), as well
as some "dual single core" systems and one "quad dual core" system.
"single dual core" => 2 slots.
"dual dual core" => 4 slots.
"quad dual core" => 8 slots.

Everything then works as you'd expect.

> 2. Is there any way to configure SGE (allocation_rule parameter?) to
> force 4 processes in the same node (2 CPUs dual-core --> 4 cores), but
> each process running into independent cores? Can we force the memory
> size for each core? The reason is that we need the same conditions on
> each core, independently where it is executed and how many processes are
> running in that node.

That's not really an SGE question, but a question for your operating system.
In general, your OS scheduler will run each active process on it's own core,
until you have more active processes that cores. Note that the operating system
itself has several generally low-activity processes that will need to steal
cycles from at least one of your cores from time to time. Whether or not you can
explicitly "lock" resources to certain processes is a question for your OS, but
usually this isn't a problem if you just set your number of slots to equal the
number of cores, then make sure you don't launch multi-threaded jobs.

In SGE, you can limit the memory -per job- to 1/4 or so of system memory using
the Limits section of the queue configuration. A job exceeding the limit will be
killed.

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