Re: ideal load threshold for cluster nodes.

On 14 Jul 2006, at 9:31 pm, Jinal Jhaveri wrote:

> Hi All,
>
> I am using load_avg=2.0 as a load threshold for our SGE nodes. I  
> avoid using the np_load_avg (load per processor) as it is mis- 
> interpreted on intel machines with hyperthreading enabled. Can  
> anyone (especialy in Bioinformatics setting running jobs like  
> blast,interpro) share their experience on what load threshold works  
> best for them?

I can't speak for SGE, but on our LSF cluster, I make the scheduling  
threshold around the number of processors - 1, and then add a tiny  
bit.  We disable hyperthreading on all of our Intel machines.

So, a 2-CPU machine has a scheduling threshold of 1.1; this means it  
should usually find itself running 2 jobs, but if it's doing anything  
else at all (cluster filesystem stuff, for example), it tends not to  
load up a second job.

This seems to work pretty well.

Tim