[PATCH v15 0/6] Add NUMA-awareness to qspinlock

[Barry Song]

On Fri, Oct 1, 2021 at 5:58 AM Waiman Long <llong at redhat.com> wrote:
>
> On 9/30/21 5:44 AM, Barry Song wrote:
> >> We have done some performance evaluation with the locktorture module
> >> as well as with several benchmarks from the will-it-scale repo.
> >> The following locktorture results are from an Oracle X5-4 server
> >> (four Intel Xeon E7-8895 v3 @ 2.60GHz sockets with 18 hyperthreaded
> >> cores each). Each number represents an average (over 25 runs) of the
> >> total number of ops (x10^7) reported at the end of each run. The
> >> standard deviation is also reported in (), and in general is about 3%
> >> from the average. The 'stock' kernel is v5.12.0,
> > I assume x5-4 server has the crossbar topology and its numa diameter is
> > 1hop, and all tests were done on this kind of symmetrical topology. Am
> > I right?
> >
> >      ┌─┐                 ┌─┐
> >      │ ├─────────────────┤ │
> >      └─┤1               1└┬┘
> >        │  1           1   │
> >        │    1       1     │
> >        │      1   1       │
> >        │        1         │
> >        │      1   1       │
> >        │     1      1     │
> >        │   1         1    │
> >       ┌┼┐1             1  ├─┐
> >       │┼┼─────────────────┤ │
> >       └─┘                 └─┘
> >
> >
> > what if the hardware is using the ring topology and other topologies with
> > 2-hops or even 3-hops such as:
> >
> >       ┌─┐                 ┌─┐
> >       │ ├─────────────────┤ │
> >       └─┤                 └┬┘
> >         │                  │
> >         │                  │
> >         │                  │
> >         │                  │
> >         │                  │
> >         │                  │
> >         │                  │
> >        ┌┤                  ├─┐
> >        │┼┬─────────────────┤ │
> >        └─┘                 └─┘
> >
> >
> > or:
> >
> >
> >      ┌───┐       ┌───┐      ┌────┐      ┌─────┐
> >      │   │       │   │      │    │      │     │
> >      │   │       │   │      │    │      │     │
> >      ├───┼───────┼───┼──────┼────┼──────┼─────┤
> >      │   │       │   │      │    │      │     │
> >      └───┘       └───┘      └────┘      └─────┘
> >
> > do we need to consider the distances of numa nodes in the secondary
> > queue? does it still make sense to treat everyone else equal in
> > secondary queue?
>
> The purpose of this patch series is to minimize cacheline transfer from
> one numa node to another. Taking the fine grained detail of the numa
> topology into account will complicate the code without much performance
> benefit from my point of view. Let's keep it simple first. We can always
> improve it later on if one can show real benefit of doing so.

for sure i am not expecting the complex  NUMA topology taken into account for
this moment. I am just curious how things will be different if topology isn't a
crossbar with 1-hop only.

when the master queue is empty, the distance of the numa node spinlock will
jump to will affect the performance. but I am not quite sure how much it will
be. just like a disk, bumping back and forth between far cylinders and sectors
might waste a lot of time.

On the other hand, some numa nodes might be very close while some others
might be very far. for example, if one socket has several DIEs, and the machine
has several sockets, cacheline coherence overhead for NUMA nodes of DIEs within
one socket might be much less than that of NUMA nodes which are in different
sockets. I assume maintaining the master/secondary queues need some
overhead especially while the system has many cores and multiple NUMA nodes,
in this case, making neighbor NUMA nodes share one master queue might win.

Anyway, we need a lot of benchmarking on this before we can really do anything
on it.  For this moment, ignoring the complicated topology should be a
better way
to start.

>
> Cheers,
> Longman

Thanks
barry