Re: [PATCH v7 00/13] fold per-CPU vmstats remotely

[Marcelo Tosatti]

On Thu, Mar 23, 2023 at 01:17:32PM +0100, Michal Hocko wrote:
> On Thu 23-03-23 07:52:22, Marcelo Tosatti wrote:
> > On Thu, Mar 23, 2023 at 08:51:14AM +0100, Michal Hocko wrote:
> > > On Wed 22-03-23 11:20:55, Marcelo Tosatti wrote:
> > > > On Wed, Mar 22, 2023 at 02:35:20PM +0100, Michal Hocko wrote:
> > > [...]
> > > > > > "Performance details for the kworker interruption:
> > > > > > 
> > > > > > oslat   1094.456862: sys_mlock(start: 7f7ed0000b60, len: 1000)
> > > > > > oslat   1094.456971: workqueue_queue_work: ... function=vmstat_update ...
> > > > > > oslat   1094.456974: sched_switch: prev_comm=oslat ... ==> next_comm=kworker/5:1 ...
> > > > > > kworker 1094.456978: sched_switch: prev_comm=kworker/5:1 ==> next_comm=oslat ...
> > > > > > 
> > > > > > The example above shows an additional 7us for the
> > > > > > 
> > > > > >         oslat -> kworker -> oslat
> > > > > > 
> > > > > > switches. In the case of a virtualized CPU, and the vmstat_update
> > > > > > interruption in the host (of a qemu-kvm vcpu), the latency penalty
> > > > > > observed in the guest is higher than 50us, violating the acceptable
> > > > > > latency threshold for certain applications."
> > > > > 
> > > > > Yes, I have seen that but it doesn't really give a wider context to
> > > > > understand why those numbers matter.
> > > > 
> > > > OK.
> > > > 
> > > > "In the case of RAN, a MAC scheduler with TTI=1ms, this causes >100us
> > > > interruption observed in a guest (which is above the safety
> > > > threshold for this application)."
> > > > 
> > > > Is that OK?
> > > 
> > > This might be a sufficient information for somebody familiar with the
> > > matter (not me). So no, not enough. We need to hear a more complete
> > > story. 
> > 
> > Michal,
> > 
> > Please refer to 
> > https://www.diva-portal.org/smash/get/diva2:541460/FULLTEXT01.pdf
> > 
> > 2.3 Channel Dependent Scheduling
> > The purpose of scheduling is to decide which terminal will transmit data on which set
> > of resource blocks with what transport format to use. The objective is to assign
> > resources to the terminal such that the quality of service (QoS) requirement is fulfilled.
> > Scheduling decision is taken every 1 ms by base station (termed as eNodeB) as the
> > same length of Transmission Time Interval (TTI) in LTE system.
> > 
> > In general:
> > 
> > https://en.wikipedia.org/wiki/Real-time_computing
> 
> Thank you, but not something I was really asking for (repeatedly). I am
> pretty aware of what RT computing is about. I am not really interested
> in a generic fluff. I am asking about specific usecases you have in mind
> when pushing these changes.
>  
> > For example, for the MAC scheduler processing must occur every 1ms,
> > and a certain amount of computation takes place (and must finish before
> > the next 1ms timeframe). A > 50us latency spike as observed by cyclictest
> > is considered a "failure".
> 
> OK, you are claiming that much but you are not really filling up other
> holes in your story. Let me just outline few questions I have. Your
> measurements talk about 7us overhead the vmstat processing might add.
> This is really far from > 50us above. 

7us in the host, for the following sched_switch events:

	oslat -> kworker
	kworker -> oslat

However, if the impact is for a virtualized application:

	oslat, executing via qemu-vcpu process in the host.

	oslat executing
	qemu-vcpu VM-EXIT
	qemu-vcpu -> kworker
	kworker -> qemu-vcpu
	qemu-vcpu VM-ENTRY

is much higher than the 7us (can be above 100us).

> You suggest that this is an effect
> of the workload running in a guest without more details. I am quite
> surprised to hear about RT expectations inside a guest system TBH.

https://www.youtube.com/watch?v=zIDwc6uDszY

> All that being said, it would be really helpful if you were more
> specific about the workload and why there is no other way but making
> vmstat infrastructure more complex (it is quite complex on its own).

The patchset is just changing vmstat_shepherd from happening locally
to happening remotely. There are a number of algorithms in the kernel
that deal with concurrent access already.

What you think this particular patchset makes things complicated
and what can be done to make it simpler?