Re: [PATCH 2/2] arm64: defconfig: Raise NR_CPUS to 256

From: Jan Glauber
Date: Mon Mar 26 2018 - 06:03:13 EST


On Mon, Mar 26, 2018 at 11:28:28AM +0200, Arnd Bergmann wrote:
> On Mon, Mar 26, 2018 at 10:52 AM, Jan Glauber
> <jan.glauber@xxxxxxxxxxxxxxxxxx> wrote:
> > On Tue, Mar 06, 2018 at 03:02:01PM +0100, Jan Glauber wrote:
> >> On Tue, Mar 06, 2018 at 02:12:29PM +0100, Arnd Bergmann wrote:
> >> > On Fri, Mar 2, 2018 at 3:37 PM, Jan Glauber <jglauber@xxxxxxxxxx> wrote:
> >> > > ThunderX1 dual socket has 96 CPUs and ThunderX2 has 224 CPUs.
> >> >
> >> > Are you sure about those numbers? From my counting, I would have expected
> >> > twice that number in both cases: 48 cores, 2 chips and 2x SMT for ThunderX
> >> > vs 52 Cores, 2 chips and 4x SMT for ThunderX2.
> >>
> >> That's what I have on those machines. I counted SMT as normal CPUs as it
> >> doesn't make a difference for the config. I've not seen SMT on ThunderX.
> >>
> >> The ThunderX2 number of 224 is already with 4x SMT (and 2 chips) but
> >> there may be other versions planned that I'm not aware of.
> >>
> >> > > Therefore raise the default number of CPUs from 64 to 256
> >> > > by adding an arm64 specific option to override the generic default.
> >> >
> >> > Regardless of what the correct numbers for your chips are, I'd like
> >> > to hear some other opinions on how high we should raise that default
> >> > limit, both in arch/arm64/Kconfig and in the defconfig file.
> >> >
> >> > As I remember it, there is a noticeable cost for taking the limit beyond
> >> > BITS_PER_LONG, both in terms of memory consumption and also
> >> > runtime performance (copying and comparing CPU masks).
> >>
> >> OK, that explains the default. My unverified assumption is that
> >> increasing the CPU masks wont be a noticable performance hit.
> >>
> >> Also, I don't think that anyone who wants performance will use
> >> defconfig. All server distributions would bump up the NR_CPUS anyway
> >> and really small systems will probably need to tune the config
> >> anyway.
> >>
> >> For me defconfig should produce a usable system, not with every last
> >> driver configured but with all the basics like CPUs, networking, etc.
> >> fully present.
> >>
> >> > I'm sure someone will keep coming up with even larger configurations
> >> > in the future, so we should try to decide how far we can take the
> >> > defaults for the moment without impacting users of the smallest
> >> > systems. Alternatively, you could add some measurements that
> >> > show how much memory and CPU time is used up on a typical
> >> > configuration for a small system (4 cores, no SMT, 512 MB RAM).
> >> > If that's low enough, we could just do it anyway.
> >>
> >> OK, I'll take a look.
> >
> > I've made some measurements on a 4 core board (Cavium 81xx) with
> > NR_CPUS set to 64 or 256:
> >
> > - vmlinux grows by 0.04 % with 256 CPUs
>
> Ok. Is this both with CONFIG_CPUMASK_OFFSTACK=n?

Yes.

> > - Kernel compile time was a bit faster with 256 CPUS (which does
> > not make sense, but at least is seems to not suffer from the change).
>
> Do you mean compiling the same kernel configuration while running
> on a system with less than 64 CPUs on either a CONFIG_NR_CPUS=64
> or CONFIG_NR_PCUS=256 kernel, or do you mean the time to compile
> a kernel with either CONFIG_NR_CPUS=64 or CONFIG_NR_CPUS=256,
> while running on the same host?

The former, compiling everything on a 4-core system using two different
kernels to compile the same thing.

> I assume the former, which is a very interesting result, possibly
> pointing to us doing something wrong in the NR_CPUS=64 case
> that could be optimized.
>
> If you ran with CONFIG_CPUMASK_OFFSTACK, that may have made
> a significant difference, but I would expect it to be faster without it.
>
> To get more insight to what is happening, you could rerun the same
> test with 'perf record' and then compare the profiles. How significant
> is the runtime difference compared to the jitter you get between normal
> runs on the same configuration?

I did retry once but the odd case that CONFIG_NR_CPUS=256 was faster
was consistent. The difference was very small though so it may be
completely due to jitter.

> > Is there a benchmark that will be better suited? Maybe even a
> > microbenchmark that will suffer from the longer cpumasks?
>
> Good question.
>
> > - Available memory decreased by 0.13% (restricted memory to 512 MB),
> > BSS increased 5.3 %
>
> 0.13% of a few hundred megabytes is still several hundred kb, right? I'd
> like to hear some other opinions on that, but it seems to be in the
> range of enabling many additional device drivers, which is something
> we don't do lightly.

Agreed, available memory was reduced by 128 KB.

--Jan