Re: Observation on NOHZ_FULL

[Andrea Righi <andrea.righi@canonical.com>]

On Tue, Jan 30, 2024 at 05:39:37AM -0800, Paul E. McKenney wrote:
> On Tue, Jan 30, 2024 at 12:27:43PM +0100, Andrea Righi wrote:
> > On Tue, Jan 30, 2024 at 12:06:49PM +0100, Uladzislau Rezki wrote:
> > > On Tue, Jan 30, 2024 at 02:17:22AM -0800, Paul E. McKenney wrote:
> > > > On Tue, Jan 30, 2024 at 07:58:18AM +0100, Andrea Righi wrote:
> > > > > Hi Joel and Paul,
> > > > > 
> > > > > comments below.
> > > > > 
> > > > > On Mon, Jan 29, 2024 at 05:16:38PM -0500, Joel Fernandes wrote:
> > > > > > Hi Paul,
> > > > > > 
> > > > > > On 1/29/2024 3:41 PM, Paul E. McKenney wrote:
> > > > > > > On Mon, Jan 29, 2024 at 05:47:39PM +0000, Joel Fernandes wrote:
> > > > > > >> Hi Guys,
> > > > > > >> Something caught my eye in [1] which a colleague pointed me to
> > > > > > >>  - CONFIG_HZ=1000 : 14866.05 bogo ops/s
> > > > > > >>  - CONFIG_HZ=1000+nohz_full : 18505.52 bogo ops/s
> > > > > > >>
> > > > > > >> The test in concern is:
> > > > > > >> stress-ng --matrix $(getconf _NPROCESSORS_ONLN) --timeout 5m --metrics-brief
> > > > > > >>
> > > > > > >> which is a CPU intensive test.
> > > > > > >>
> > > > > > >> Any thoughts on what else can attribute a 30% performance increase
> > > > > > >> versus non-nohz_full ? (Confession: No idea if the baseline is
> > > > > > >> nohz_idle or no nohz at all). If it is 30%, I may want to evaluate
> > > > > > >> nohz_full on some of our limited-CPU devices :)
> > > > > > > 
> > > > > > > The usual questions.  ;-)
> > > > > > > 
> > > > > > > Is this repeatable?  Is it under the same conditions of temperature,
> > > > > > > load, and so on?  Was it running on bare metal or on a guest OS?  If on a
> > > > > > > guest OS, what was the load from other guest OSes on the same hypervisor
> > > > > > > or on the hypervisor itself?
> > > > > 
> > > > > That was the result of a quick test, so I expect it has some fuzzyness
> > > > > in there.
> > > > > 
> > > > > It's an average of 10 runs, it was bare metal (my laptop, 8 cores 11th
> > > > > Gen Intel(R) Core(TM) i7-1195G7 @ 2.90GHz), *but* I wanted to run the
> > > > > test with the default Ubuntu settings, that means having "power mode:
> > > > > balanced" enabled. I don't know exactly what it's doing (I'll check how
> > > > > it works in details), I think it's using intel p-states IIRC.
> > > > > 
> > > > > Also, the system was not completely isolated (my email client was
> > > > > running) but the system was mostly idle in general.
> > > > > 
> > > > > I was already planning to repeat the tests in a more "isolated"
> > > > > environment and add details to the bug tracker.
> > > > > 
> > > > > > > 
> > > > > > > The bug report ad "CONFIG_HZ=250 : 17415.60 bogo ops/s", which makes
> > > > > > > me wonder if someone enabled some heavy debug that is greatly
> > > > > > > increasing the overhead of the scheduling-clock interrupt.
> > > > > > > 
> > > > > > > Now, if that was the case, I would expect the 250HZ number to have
> > > > > > > three-quarters of the improvement of the nohz_full number compared
> > > > > > > to the 1000HZ number:
> > > > > > >> 17415.60-14866.05=2549.55
> > > > > > > 18505.52-14866.05=3639.47
> > > > > > > 
> > > > > > > 2549.55/3639.47=0.70
> > > > > > 
> > > > > > I wonder if the difference here could possibly also be because of CPU idle
> > > > > > governor. It may behave differently at differently clock rates so perhaps has
> > > > > > different overhead.
> > > > > 
> > > > > Could be, but, again, the balanced power mode could play a major role
> > > > > here.
> > > > > 
> > > > > > 
> > > > > > I have added trying nohz full to my list as well to evaluate. FWIW, when we
> > > > > > moved from 250HZ to 1000HZ, it actually improved power because the CPUidle
> > > > > > governor could put the CPUs in deeper idle states more quickly!
> > > > > 
> > > > > Interesting, another benefit to add to my proposal. :)
> > > > > 
> > > > > > 
> > > > > > > OK, 0.70 is not *that* far off of 0.75.  So what debugging does that
> > > > > > > test have enabled?  Also, if you use tracing (or whatever) to measure
> > > > > > > the typical duration of the scheduling-clock interrupt and related things
> > > > > > > like softirq handlers, does it fit with these numbers?  Such a measurment
> > > > > > > would look at how long it took to get back into userspace.
> > > > 
> > > > Just to emphasize...
> > > > 
> > > > The above calculations show that your measurements are close to what you
> > > > would expect if scheduling-clock interrupts took longer than one would
> > > > expect.  Here "scheduling-clock interrupts" includes softirq processing
> > > > (timers, networking, RCU, ...)  that piggybacks on each such interrupt.
> > > > 
> > > > Although softirq makes the most sense given the amount of time that must
> > > > be consumed, for the most part softirq work is conserved.  which suggests
> > > > that you should also at the rest of the system to check whether the
> > > > reported speedup is instead due to this work simply being moved to some
> > > > other CPU.
> > > > 
> > > > But maybe the fat softirqs are due to some debugging option that Ubuntu
> > > > enabled.  In which case checking up on the actual duration (perhaps
> > > > using some form of tracing) would provide useful information.  ;-)
> > > > 
> > > As a first step i would have a look at perf figures what is going on
> > > during a test run. For such purpose the "perf" tool can be used. As a
> > > basic step it can be run in a "top" mode:
> > > 
> > > perf top -a -g -e cycles:k 
> > > 
> > > Sorry for the noise :)
> > 
> > Yep, I'm planning to do better tests and collect more info (perf,
> > bpftrace). Also making sure that we don't have some crazy debugging
> > config enabled in the Ubuntu kernel, as correctly pointed by Paul. But
> > first of all I need to repeat the tests in a more isolated environment,
> > just to make sure we're looking at reasonable numbers here.
> 
> Very good, and looking forward to seeing what you come up with!
> 
> 							Thanx, Paul

I repeated some tests in a more isolated environment and posted the
results here:
https://discourse.ubuntu.com/t/enable-low-latency-features-in-the-generic-ubuntu-kernel-for-24-04/42255

Highlights:

 - stress-ng --matrix seems quite unpredictable to be used a benchmarks
   in this scenario (the bogo-ops/s are very susceptible to any kind of
   interference, so even if in the long runs NO_HZ_FULL still seems to
   provide some benefits looking at the average, we also need to
   consider that there might be a significant error in the measurements,
   standard deviation was pretty high)

 - fio doing short writes (in page cache) seems to perform like 2x
   better in terms of iops with nohz_full, respect to the other cases
   and it performs 2x slower with large IO writes (not sure why... need
   to investigate more)

 - with lazy RCU enabled hrtimer_interrupt() takes like 2x more to
   return, respect to the other cases (is this expected?)

Let me know if you notice anything odd, if you need more information,
etc.

Thanks,
-Andrea