Re: [PATCH RFC 3/8] sched/ext: Add a DL server for sched_ext tasks

[Peter Zijlstra]

On Sat, Mar 15, 2025 at 07:15:27PM -0400, Joel Fernandes wrote:
> 
> 
> On 3/15/2025 3:22 AM, Peter Zijlstra wrote:
> > On Fri, Mar 14, 2025 at 10:21:50PM -0400, Joel Fernandes wrote:
> >> sched_ext currently suffers starvation due to RT. The same workload when
> >> converted to EXT can get zero runtime if RT is 100% running, causing EXT
> >> processes to stall. Fix it by adding a DL server for EXT.
> > 
> > This needs a lot more words on why you need a second server. Because I
> > don't think you do.
> 
> Sure, I will add more words to the change log to explain rationale. When you say
> "I don't think you do", do you mean that both FAIR and EXT could be served by
> the same server? 

Yeah, because now you get two deadline entities both having a
reservation on bandwidth. One of which is not going to be used -- this
is not nice.

> If so, that will not handle the case where the system has both
> FAIR and EXT tasks in the mix (EXT has a partial mode where certain tasks can be
> made EXT with certain others left as FAIR) and FAIR runs 100% and starves EXT.

Well, you did not mention that issue, you only babbled about RT.

I did point out that issue with ext, and TJ said this mixed mode wasn't
really meant to be used or somesuch.

So if that's changed, this needs a separate discussion.

Also; I gotta ask, why is nvidia looking at ext ?