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

[Marcelo Tosatti <mtosatti@redhat.com>]

On Wed, Mar 22, 2023 at 02:35:20PM +0100, Michal Hocko wrote:
> On Wed 22-03-23 08:23:21, Marcelo Tosatti wrote:
> > On Wed, Mar 22, 2023 at 11:13:02AM +0100, Michal Hocko wrote:
> > > On Mon 20-03-23 16:07:29, Marcelo Tosatti wrote:
> > > > On Mon, Mar 20, 2023 at 07:25:55PM +0100, Michal Hocko wrote:
> > > > > On Mon 20-03-23 15:03:32, Marcelo Tosatti wrote:
> > > > > > This patch series addresses the following two problems:
> > > > > > 
> > > > > > 1. A customer provided evidence indicating that a process
> > > > > >    was stalled in direct reclaim:
> > > > > > 
> > > > > This is addressed by the trivial patch 1.
> > > > > 
> > > > > [...]
> > > > > >  2. With a task that busy loops on a given CPU,
> > > > > >     the kworker interruption to execute vmstat_update
> > > > > >     is undesired and may exceed latency thresholds
> > > > > >     for certain applications.
> > > > > 
> > > > > Yes it can but why does that matter?
> > > > 
> > > > It matters for the application that is executing and expects
> > > > not to be interrupted.
> > > 
> > > Those workloads shouldn't enter the kernel in the first place, no?
> > 
> > It depends on the latency requirements and individual system calls.
> > 
> > > Otherwise the in kernel execution with all the direct or indirect
> > > dependencies (e.g. via locks) can throw any latency expectations off the
> > > window.
> > > 
> > > > > > By having vmstat_shepherd flush the per-CPU counters to the
> > > > > > global counters from remote CPUs.
> > > > > > 
> > > > > > This is done using cmpxchg to manipulate the counters,
> > > > > > both CPU locally (via the account functions),
> > > > > > and remotely (via cpu_vm_stats_fold).
> > > > > > 
> > > > > > Thanks to Aaron Tomlin for diagnosing issue 1 and writing
> > > > > > the initial patch series.
> > > > > > 
> > > > > > 
> > > > > > 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.
> > > > > 
> > > > > I do not think we have ever promissed any specific latency guarantees
> > > > > for vmstat. These are statistics have been mostly used for debugging
> > > > > purposes AFAIK. I am not aware of any specific user space use case that
> > > > > would be latency sensitive. Your changelog doesn't go into details there
> > > > > either.
> > > > 
> > > > There is a class of workloads for which response time can be
> > > > of interest. MAC scheduler is an example:
> > > > 
> > > > https://par.nsf.gov/servlets/purl/10090368
> > > 
> > > Yes, I am not disputing low latency workloads in general. I am just
> > > saying that you haven't really established a very sound justification
> > > here.
> > 
> > The -v7 cover letter was updated with additional details, 
> > as you requested (perhaps you missed it):
> > 
> > "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?


> > > Of course there are workloads which do not want to conflict with
> > > any in kernel house keeping. Those have to be configured and implemented
> > > very carefully though. Vmstat as such should not collide with those
> > > workloads as long as they do not interact with the kernel in a way
> > > counters are updated. Is this hard or impossible to avoid? 
> > 
> > The practical problem we have been seeing is -RT app initialization.
> > For example:
> > 
> > 1) mlock();
> > 2) enter loop without system calls
> 
> OK, that is what I have kinda expected. Would have been better to
> mention it explicitly.
> 
> I expect this to be a very common pattern and vmstat might not be the
> only subsystem that could interfere later on. Would it make more sense
> to address this by a more generic solution? E.g. a syscall to flush all
> per-cpu caches so they won't interfere later unless userspace hits the
> kernel path in some way (e.g. flush_cpu_caches(cpu_set_t cpumask, int flags)?
> The above pattern could then be implemented as
> 
> 	do_initial_setup()
> 	sched_setaffinity(getpid(), cpumask);
> 	flush_cpu_caches(cpumask, 0);
> 	do_userspace_loop()

I would argue that fixing this without introducing a userspace tunable 
is more generic as all programs (modified to use a syscall or not)
benefit from the improvement. HPC workloads, for example.

But it might be necessary to do what you suggest for other
reasons (where you'd want a behaviour to be enabled which
is undesired for other application types).