AWS Thinkbox Discussion Forums

Secondary pool misunderstanding

What I’m trying to accomplish is have specific rendernodes pick up jobs first, then spill over to other rendernodes, but that’s not happening.

An example:
rendernode_1: 3d,all
rendernode_2: 3d,all
rendernode_3: all
rendernode_4: all

If I submit a 3d job, rendernodes 1 and 2 should get tasks first, then rendernodes 3 and 4 if there are more tasks. What’s happening is slaves are getting tasks in a random order, regardless of pool. What am I doing wrong here?

In Deadline, Slaves decide which Job’s to pick up based on the repository Job Scheduling settings. A side effect of this is that the order in which Slaves pick up a Job will be random.

The Secondary Pool setting in the Job Properties is also evaluated from the Slave’s perspective. The Slave will first scan all Jobs looking for a match to the primary pool. If no eligible Jobs are identified, it will then re-scan all Jobs looking for a match to the secondary pool, if assigned. In short, the Secondary Pool setting lets the Slave consider a wider (secondary) set of Jobs if the normal (primary) set of Jobs comes up empty. However, even in this case, the order in which Slaves pick up Jobs is random.

In short, I don’t think there’s a way (out of the box) to ensure that all Slaves in a Job’s primary pool are saturated before other Slaves start to participate on the Job. That said, Deadline has become so scriptable that there’s usually a way to accomplish a goal. For example, with clever use of machine white lists and scripting, you could heavily influence the order in which Slaves pick up a Job, but I think this would ultimately be more trouble than it’s worth.

Ok, thanks, I understand now.

My ultimate goal is to have jobs land on all the newer drones before spilling over to the older, slower drones. Maybe I need to rethink the pool/group structure.

We have had requests for faster machines to be more likely to pick up Tasks for a Job than slower machines, so this is already on our Job scheduling wishlist. It is especially annoying when the last few Tasks of a Job get picked up by really slow machines (what I call the “Slow Tail effect”).

Yes, I experience that all the time. This year we are also using workstations for renders, and it would be nice to prioritize drones over workstations.

Privacy | Site terms | Cookie preferences