starion83
Nov. 26, 2020 19:41:18
Hi everyone,
I'm having a bit of trouble working out the correct setup to utilize all threads on all machines through deadline for something like outputting from a ROPgeo to disk.
I can distribute a job across these machines, however the concurrent tasks are limited to 16 in deadline scheduler and this does not take advantage of all cores.
Does anyone have an example of a setup that would achieve this? I'm running H18.
Many thanks in advance!
starion83
Nov. 28, 2020 22:52:02
I've explored using wedges a little but without any luck getting full usage of all CPU cores over deadline. Anyone?
chrisgreb
Nov. 30, 2020 09:20:08
You could try running more clients on each machine?
starion83
Nov. 30, 2020 17:51:33
Hey Chris, thanks for your reply!
I think because this would mean I'd have to run a machine for every thread, it's not an ideal solution is I'd have to start up many clients. Ideally I'd like to just run each client with the same capacity as I would if I was running the PDG locally where it can max out the CPU.
chrisgreb
Nov. 30, 2020 18:03:53
Did you try changing the Job Options/ Concurrent Tasks parm on the Deadline Scheduler TOP?
martinkindl83
Nov. 30, 2020 18:27:59
chrisgreb
Concurren
I believe concurrent tasks are limited to 16only by deadline
chrisgreb
Nov. 30, 2020 19:09:22
I see, you would need 4 workers to saturate a 64 core machine. This sounds like something that would need to be taken up with Thinkbox.
One work around may be to break up the graph into multiple graphs which use Local Scheduler. Then a top-level graph would use Top Fetch nodes with Deadline scheduler to process those sub graphs in parallel, so each smaller graph would consume all the resources of the machine they were assigned to. However, this would no doubt obfuscate your graph.
starion83
Dec. 30, 2020 01:55:05
chrisgreb
I see, you would need 4 workers to saturate a 64 core machine. This sounds like something that would need to be taken up with Thinkbox.
One work around may be to break up the graph into multiple graphs which use Local Scheduler. Then a top-level graph would use Top Fetch nodes with Deadline scheduler to process those sub graphs in parallel, so each smaller graph would consume all the resources of the machine they were assigned to. However, this would no doubt obfuscate your graph.
Thanks for the reply, and sorry for the delay with mine!
Once I get a chance I'll look into this again.