EricSheng
Sept. 9, 2019 12:01:21
It seems like the Invoke Node isn't limited by the Local Scheduler, no matter how many CPU counts I set in the Local Scheduler Node, the Invoke Node will use more CPU. And in my case, I run some OpenCL code in Invoke Node, and it created 15 workitems parallelly and cause the GPU crashing. And then my computer just restarts by itself.
So is there any way to limit the CPU use for Invoke Node? Or how can I run the Invoke node not parallelly? Thanks.
chrisgreb
Sept. 9, 2019 12:09:31
There's an existing RFE for this (99118).
In the meantime you could consider running your invoke in a For Each Loop, which would ensure each invoke iteration runs in series.
EricSheng
Sept. 9, 2019 21:14:26
chrisgreb
There's an existing RFE for this (99118).
In the meantime you could consider running your invoke in a For Each Loop, which would ensure each invoke iteration runs in series.
Thanks, Chrisgreb:
I've tried to use For Each Loop, but it seems like still use too many threads, as you can see in the GIF, I've limited only use 3 CPUs, but the For Each Loop still generate 7 work_item parallelly.
chrisgreb
Sept. 9, 2019 21:34:15
Try toggling the “Iterations from upstream items” parm on the feedback begin block
https://www.sidefx.com/docs/houdini/tops/looping.html [
www.sidefx.com]
EricSheng
Sept. 10, 2019 04:13:13
chrisgreb
Try toggling the “Iterations from upstream items” parm on the feedback begin block
https://www.sidefx.com/docs/houdini/tops/looping.html [www.sidefx.com]
Thank you for your reply, Chris:
Yes, it works after toggling the “Iterations from upstream items” parm but unfortunately, it's not compatible with dynamic work_items While I'm dealing with dynamic work_items.
EricSheng
Nov. 11, 2019 02:26:00
chrisgreb
There's an existing RFE for this (99118).
In the meantime you could consider running your invoke in a For Each Loop, which would ensure each invoke iteration runs in series.
Hi, any progress yet? Will this feature be available in H18? Thanks.
chrisgreb
Nov. 13, 2019 11:10:52
EricSheng
chrisgreb
There's an existing RFE for this (99118).
In the meantime you could consider running your invoke in a For Each Loop, which would ensure each invoke iteration runs in series.
Hi, any progress yet? Will this feature be available in H18? Thanks.
No it's sitting in the backlog and won't make it to 18.