Re: [PATCH] sched/fair: no sync wakeup from interrupt context

[Libo Chen]

On 7/13/22 13:51, Tim Chen wrote:
> On Wed, 2022-07-13 at 12:17 -0700, Libo Chen wrote:
> > On 7/13/22 09:40, Tim Chen wrote:
> > > On Mon, 2022-07-11 at 15:47 -0700, Libo Chen wrote:
> > > > Barry Song first pointed out that replacing sync wakeup with regular wakeup
> > > > seems to reduce overeager wakeup pulling and shows noticeable performance
> > > > improvement.[1]
> > > >
> > > > This patch argues that allowing sync for wakeups from interrupt context
> > > > is a bug and fixing it can improve performance even when irq/softirq is
> > > > evenly spread out.
> > > >
> > > > For wakeups from ISR, the waking CPU is just the CPU of ISR and the so-called
> > > > waker can be any random task that happens to be running on that CPU when the
> > > > interrupt comes in. This is completely different from other wakups where the
> > > > task running on the waking CPU is the actual waker. For example, two tasks
> > > > communicate through a pipe or mutiple tasks access the same critical section,
> > > > etc. This difference is important because with sync we assume the waker will
> > > > get off the runqueue and go to sleep immedately after the wakeup. The
> > > > assumption is built into wake_affine() where it discounts the waker's presence
> > > > from the runqueue when sync is true. The random waker from interrupts bears no
> > > > relation to the wakee and don't usually go to sleep immediately afterwards
> > > > unless wakeup granularity is reached. Plus the scheduler no longer enforces the
> > > > preepmtion of waker for sync wakeup as it used to before
> > > > patch f2e74eeac03ffb7 ("sched: Remove WAKEUP_SYNC feature"). Enforcing sync
> > > > wakeup preemption for wakeups from interrupt contexts doesn't seem to be
> > > > appropriate too but at least sync wakeup will do what it's supposed to do.
> > > Will there be scenarios where you do want the task being woken up be pulled
> > > to the CPU where the interrupt happened, as the data that needs to be accessed is
> > > on local CPU/NUMA that interrupt happened?  For example, interrupt associated with network
> > > packets received.  Sync still seems desirable, at least if there is no task currently
> > > running on the CPU where interrupt happened.  So maybe we should have some consideration
> > > of the load on the CPU/NUMA before deciding whether we should do sync wake for such
> > > interrupt.
> >   There are only two places where sync wakeup matters: wake_affine_idle() and wake_affine_weight().
> > In wake_affine_idle(), it considers pulling if there is one runnable on the waking CPU because
> > of the belief that this runnable will voluntarily get off the runqueue. In wake_affine_weight(),
> > it basically takes off the waker's load again assuming the waker goes to sleep after the wakeup.
> > My argument is that this assumption doesn't really hold for wakeups from the interrupt contexts
> > when the waking CPU is non-idle. Wakeups from task context? sure, it seems to be a reasonable
> > assumption.
> I agree with you that the the sync case load computation for wake_affine_idle()
> and wake_affine_weight() is incorrect when waking a task from the interrupt context.
> In this light, your proposal makes sense.
>
> > For your idle case, I totally agree but I don't think having sync or not will actually
> > have any impacts here giving what the code does. Real impact comes from Mel's patch 7332dec055f2457c3
> > which makes it less likely to pull tasks when the waking CPU is idle. I believe we should consider
> > reverting 7332dec055f2 because a significant RDS latency regression has been spotted recently on our
> > system due to this patch.
> The commit 7332dec055f2 prevented cross NUMA node pulling.  I think if the
> waking CPU's NUMA node's average load is less than the prev CPU's NUMA node,
> this cross NUMA node pull could be allowed for better load distribution.
Yeah, we should rewrite wake_affine_weight() so that it compares average 
loads of two nodes
instead of two rq loads.


Libo

> > > Can you provide some further insights on why pgebench is helped at low load
> > > case?  Is it because the woken tasks tend to stay put and not get moved around with interrupts
> > > and maintain cache warmth?
> >   Yes, and for read-only database workloads, the cache (whether it's incoming packet or not) on the interrupt
> > CPU isn't as performance critical as cache from its previous CPU where the db task run to process data.
> > To give you an example, consider a db client/server case, a client sends a request for a select query
> > through the network, the server accepts the query request and does all the heavy lifting and sends the result
> > back. For the server, the incoming packet is just a line of query whereas the CPU and its L3 db process previously
> > on has all the warm db caches, pulling it away from them is a crime :)  This may seem to be a little contradictory
> > to what I said earlier about the idle case and Mel's patch, but ¯\_(ツ)_/¯ it's hard to make all the workloads out
> > there happy. Anyway like I said earlier, this patch doesn't affect the idle case
> >
> > At higher load, sync in wake_affine_idle() doesn't really matter because the waking CPU could easily have more than
> > 1 runnable tasks. Sync in wake_affine_weight() also doesn't matter much as both sides have work to do, and cache
> > benefit of not pulling decreases simply because there are a lot more db processes under the same L3, they can compete
> > for the same cachelines.
> >
> > Hope my explanation helps!
> Yes, that makes sense.
>
> Tim