Re: [PATCH 00/40] Memory allocation profiling

[Roman Gushchin <roman.gushchin@linux.dev>]

On Mon, May 01, 2023 at 11:08:05AM -0700, Suren Baghdasaryan wrote:
> On Mon, May 1, 2023 at 10:47 AM Roman Gushchin <roman.gushchin@linux.dev> wrote:
> >
> > On Mon, May 01, 2023 at 09:54:10AM -0700, Suren Baghdasaryan wrote:
> > > Performance overhead:
> > > To evaluate performance we implemented an in-kernel test executing
> > > multiple get_free_page/free_page and kmalloc/kfree calls with allocation
> > > sizes growing from 8 to 240 bytes with CPU frequency set to max and CPU
> > > affinity set to a specific CPU to minimize the noise. Below is performance
> > > comparison between the baseline kernel, profiling when enabled, profiling
> > > when disabled (nomem_profiling=y) and (for comparison purposes) baseline
> > > with CONFIG_MEMCG_KMEM enabled and allocations using __GFP_ACCOUNT:
> > >
> > >                       kmalloc                 pgalloc
> > > Baseline (6.3-rc7)    9.200s                  31.050s
> > > profiling disabled    9.800 (+6.52%)          32.600 (+4.99%)
> > > profiling enabled     12.500 (+35.87%)        39.010 (+25.60%)
> > > memcg_kmem enabled    41.400 (+350.00%)       70.600 (+127.38%)
> >
> > Hm, this makes me think we have a regression with memcg_kmem in one of
> > the recent releases. When I measured it a couple of years ago, the overhead
> > was definitely within 100%.
> >
> > Do you understand what makes the your profiling drastically faster than kmem?
> 
> I haven't profiled or looked into kmem overhead closely but I can do
> that. I just wanted to see how the overhead compares with the existing
> accounting mechanisms.

It's a good idea and I generally think that +25-35% for kmalloc/pgalloc
should be ok for the production use, which is great!
In the reality, most workloads are not that sensitive to the speed of
memory allocation.

> 
> For kmalloc, the overhead is low because after we create the vector of
> slab_ext objects (which is the same as what memcg_kmem does), memory
> profiling just increments a lazy counter (which in many cases would be
> a per-cpu counter).

So does kmem (this is why I'm somewhat surprised by the difference).

> memcg_kmem operates on cgroup hierarchy with
> additional overhead associated with that. I'm guessing that's the
> reason for the big difference between these mechanisms but, I didn't
> look into the details to understand memcg_kmem performance.

I suspect recent rt-related changes and also the wide usage of
rcu primitives in the kmem code. I'll try to look closer as well.

Thanks!