[RFC] Make use of non-dynamic dmabuf in RDMA

[Oded Gabbay]

On Wed, Aug 25, 2021 at 6:14 PM Christian König
<christian.koenig at amd.com> wrote:
>
> Am 25.08.21 um 16:47 schrieb Jason Gunthorpe:
> > On Wed, Aug 25, 2021 at 03:51:14PM +0200, Christian König wrote:
> >> Am 25.08.21 um 14:38 schrieb Jason Gunthorpe:
> >>> On Wed, Aug 25, 2021 at 02:27:08PM +0200, Christian König wrote:
> >>>> Am 25.08.21 um 14:18 schrieb Jason Gunthorpe:
> >>>>> On Wed, Aug 25, 2021 at 08:17:51AM +0200, Christian König wrote:
> >>>>>
> >>>>>> The only real option where you could do P2P with buffer pinning are those
> >>>>>> compute boards where we know that everything is always accessible to
> >>>>>> everybody and we will never need to migrate anything. But even then you want
> >>>>>> some mechanism like cgroups to take care of limiting this. Otherwise any
> >>>>>> runaway process can bring down your whole system.
> >>>>> Why? It is not the pin that is the problem, it was allocating GPU
> >>>>> dedicated memory in the first place. pinning it just changes the
> >>>>> sequence to free it. No different than CPU memory.
> >>>> Pinning makes the memory un-evictable.
> >>>>
> >>>> In other words as long as we don't pin anything we can support as many
> >>>> processes as we want until we run out of swap space. Swapping sucks badly
> >>>> because your applications become pretty much unuseable, but you can easily
> >>>> recover from it by killing some process.
> >>>>
> >>>> With pinning on the other hand somebody sooner or later receives an -ENOMEM
> >>>> or -ENOSPC and there is no guarantee that this goes to the right process.
> >>> It is not really different - you have the same failure mode once the
> >>> system runs out of swap.
> >>>
> >>> This is really the kernel side trying to push a policy to the user
> >>> side that the user side doesn't want..
> >> But which is still the right thing to do as far as I can see. See userspace
> >> also doesn't want proper process isolation since it takes extra time.
> > Why? You are pushing a policy of resource allocation/usage which
> > more properly belongs in userspace.
> >
> >> Kernel development is driven by exposing the hardware functionality in a
> >> save and manageable manner to userspace, and not by fulfilling userspace
> >> requirements.
> > I don't agree with this, that is a 1980's view of OS design. So much
> > these days in the kernel is driven entirely by boutique userspace
> > requirements and is very much not about the classical abstract role of
> > an OS.
>
> But it's still true never the less. Otherwise you would have libraries
> for filesystem accesses and no system security to speak of.
>
> >>> Dedicated systems are a significant use case here and should be
> >>> supported, even if the same solution wouldn't be applicable to someone
> >>> running a desktop.
> >> And exactly that approach is not acceptable.
> > We have knobs and settings all over the place to allow Linux to
> > support a broad set of use cases from Android to servers, to HPC. So
> > long as they can co-exist and the various optional modes do not
> > critically undermine the security of the kernel, it is well in line
> > with how things have been evolving in the last 15 years.
>
> Yeah, that's exactly what I'm talking about by adding cgroup or similar.
> You need a knob to control this.
>
> > Here you are talking entirely about policy to control memory
> > allocation, which is already well trodden ground for CPU memory.
> >
> > There are now endless boutique ways to deal with this, it is a very
> > narrow view to say that GPU memory is so special and different that
> > only one way can be the correct/allowed way.
>
> Well I'm not talking about GPU memory in particular here. This is
> mandatory for any memory or saying more general any resource.
>
> E.g. you are not allowed to pin large amount of system memory on a
> default installation for exactly those reasons as well.

Unless you are working on a h/w architecture that is designed
specifically for a single user.
At least in our h/w architecture, we aim for 100% utilization of the
h/w when you are running DL training workloads, as opposed to what is
happening inside a GPU when you are running training.
Therefore, it is counter-productive to serve multiple users from a
performance perspective.
In fact, the training problem is so large, that a single ASIC is
hardly sufficient, and you need to run on anywhere between 8 ASICs to
64-128 ASICs to get a reasonable time for training.
So there is no point in serving multiple users in that case and that's
why our device memory (HBM) is always pinned.

This is totally different from the usual role of a GPU, where you
serve the desktop and multiple other applications that draw on the
screen.

I wouldn't want some knob in the kernel to control what is the limit
of memory I can pin or not. I just don't think it is useful and/or
user friendly.

btw, regarding ODP, we don't support it in h/w (yet) because 99% of
the use-cases in DL training will suffer from performance issues if we
don't pin the host memory.
There are extreme topologies, with 10TB of parameters that may require
this, but like I said, it's very rare and not worth the effort.

Thanks,
Oded

>
> That you can have a knob to disable this behavior for your HPC system is
> perfectly fine, but I thing what Dave notes here as well that this is
> most likely not the default behavior we want.
>
> Christian.
>
> >
> > Jason
>