Hi, In GPU scheduling with timeslice, there is a formula of how to calculate worst-case latency(high). What’s strange is that the two pictures above and below don’t correspond.
So is the formula universal or just for the example of this picture?
Details is in Runlist Interleave Frequency | NVIDIA Docs
Dear @gulatuoba.cn ,
Just want to confirm if you are using DRIVE or Jetson platform?
Yes， Both have used
Dear @gulatuoba.cn ,
The formula is not generic. In the first diagram, it has one high, one medium and two low priority applications so high priority application gets more time slices(like 4 times than low) where as in second diagram it has all high priority aplications and one low priority. So Hn gets a chance after H1,… Hn-1 finishes and in worst case the low priority application gets a chance to schedule multiple times and it gets finished . So the low priority application execution time is added.
Thank you， I got it
Another question 1:
From the picture, I don’t see the context-switch overhead between H1 and H2, H2 and H3, are they ignored?
question 2:
In the document, are the H1 and H2 different processes or different cuda Context task?
Thank you!
Hi @gulatuoba.cn,
For you information, The run list contains list of channel ID and each channel Id is corresponding to a context. Each application gets executed in a context. Now, each channel will have time slice and when the time slice expires the next available channel needs to be executed. So it causes the channel reset to kick in. Now if we want to give more priority to an application, we have repeat the corresponding channel many times in the run list to allow the application to get more number of chances(time slices).
A context switch time out occur when an application can not finished task with in time slice. This causes the channel to reset to allow the next context to be loaded. If an application gets finished with in time slice, then no need for context switch time out or channel reset
I think, in the diagram H1,H2,H3 running in same context.
Thank you reply!
About question 2, I still have some questions. I have done experiments, and theoretically the minimum granularity for setting timeslice is tsgid (I don’t know what it is specifically), but currently the minimum granularity I can set for timeslice is context. So if H1, H2, and H3 are the same context, how to set timeslice for them respectively?
