[Clusterusers] Changes to cluster for more fair usage
Jaime Davila
jjdCCS at hampshire.edu
Fri Jun 28 14:55:38 EDT 2013
Believe it or not, I can see some difference already. Thanks!
On 06/28/2013 02:52 PM, Wm. Josiah Erikson wrote:
> So, I've made the following changes:
>
> 1. I've removed the "tom" tag from compute-4-2 through compute-4-4,
> since he cannot use it as well as others can, and this will leave a
> large number of slots always open for shorter-running tasks
> 2. I've increased the number of slots on compute-4-2 through
> compute-4-5 to half the number of cores they have. I tried going
> higher, but weird things started happening with thread creation (I had
> to restart one of your jobs because of it, Jaime - if you were
> wondering why that happened, that's why). This will also enable Jaime
> to launch 100 concurrent tasks, which is best for his runs. Bassam
> should also, at the same time, be able to take up one slot per node,
> which should provide maximum usage of Rack 4. Blender is currently set
> to only launch one task per blade, and we will probably keep it that way.
> 3. I've changed our dispatching scheme to:
>
> "P+ATCL" -- Active Task Count Leveling, this mode also first
> # sorts jobs by strict priority, then within a group of jobs with
> # the same priority it prefers to assign available blades to jobs
> # with the fewest active tasks. This mode tends to allocate roughly
> # the same number of blades to each job, while favoring older jobs
> # over newly spooled ones. Given roughly equal numbers of blades,
> # jobs with short-running tasks will finish sooner than jobs with
> # long-running tasks, under this scheme.
>
> I was wrong, by the way - we were using P+RR before.
>
> Once Tom's jobs that are currently on Rack 4 finish, we should start
> seeing more fair usage, I think.
>
> We may need to tweak this again if/when another use-case comes along.
>
> All the best,
>
--
******************************************************
Jaime J. Dávila
Associate Professor of Computer Science
Hampshire College
jdavila at hampshire dot edu
http://helios.hampshire.edu/jdavila
*******************************************************
More information about the Clusterusers
mailing list