Re: [dpdk-dev] [PATCH v3 0/9] GPU library

[Elena Agostini <eagostini@nvidia.com>]

From: Jerin Jacob <jerinjacobk@gmail.com>
Date: Monday, 11 October 2021 at 15:30
To: NBU-Contact-Thomas Monjalon <thomas@monjalon.net>
Cc: techboard@dpdk.org <techboard@dpdk.org>, Elena Agostini <eagostini@nvidia.com>, dpdk-dev <dev@dpdk.org>
Subject: Re: [dpdk-dev] [PATCH v3 0/9] GPU library
> On Mon, Oct 11, 2021 at 6:14 PM Thomas Monjalon <thomas@monjalon.net> wrote:
> >
> > 11/10/2021 13:41, Jerin Jacob:
> > > On Mon, Oct 11, 2021 at 3:57 PM Thomas Monjalon <thomas@monjalon.net> wrote:
> > > > 11/10/2021 11:29, Jerin Jacob:
> > > > > On Mon, Oct 11, 2021 at 2:42 PM Thomas Monjalon <thomas@monjalon.net> wrote:
> > > > > > 11/10/2021 10:43, Jerin Jacob:
> > > > > > > On Mon, Oct 11, 2021 at 1:48 PM Thomas Monjalon <thomas@monjalon.net> wrote:
> > > > > > > > 10/10/2021 12:16, Jerin Jacob:
> > > > > > > > > On Fri, Oct 8, 2021 at 11:13 PM <eagostini@nvidia.com> wrote:
> > > > > > > > > >
> > > > > > > > > > From: eagostini <eagostini@nvidia.com>
> > > > > > > > > >
> > > > > > > > > > In heterogeneous computing system, processing is not only in the CPU.
> > > > > > > > > > Some tasks can be delegated to devices working in parallel.
> > > > > > > > > >
> > > > > > > > > > The goal of this new library is to enhance the collaboration between
> > > > > > > > > > DPDK, that's primarily a CPU framework, and GPU devices.
> > > > > > > > > >
> > > > > > > > > > When mixing network activity with task processing on a non-CPU device,
> > > > > > > > > > there may be the need to put in communication the CPU with the device
> > > > > > > > > > in order to manage the memory, synchronize operations, exchange info, etc..
> > > > > > > > > >
> > > > > > > > > > This library provides a number of new features:
> > > > > > > > > > - Interoperability with GPU-specific library with generic handlers
> > > > > > > > > > - Possibility to allocate and free memory on the GPU
> > > > > > > > > > - Possibility to allocate and free memory on the CPU but visible from the GPU
> > > > > > > > > > - Communication functions to enhance the dialog between the CPU and the GPU
> > > > > > > > >
> > > > > > > > > In the RFC thread, There was one outstanding non technical issues on this,
> > > > > > > > >
> > > > > > > > > i.e
> > > > > > > > > The above features are driver specific details. Does the DPDK
> > > > > > > > > _application_ need to be aware of this?
> > > > > > > >
> > > > > > > > I don't see these features as driver-specific.
> > > > > > >
> > > > > > > That is the disconnect. I see this as more driver-specific details
> > > > > > > which are not required to implement an "application" facing API.
> > > > > >
> > > > > > Indeed this is the disconnect.
> > > > > > I already answered but it seems you don't accept the answer.
> > > > >
> > > > > Same with you. That is why I requested, we need to get opinions from others.
> > > > > Some of them already provided opinions in RFC.
> > > >
> > > > This is why I Cc'ed techboard.
> > >
> > > Yes. Indeed.
> > >
> > > >
> > > > > > First, this is not driver-specific. It is a low-level API.
> > > > >
> > > > > What is the difference between low-level API and driver-level API.
> > > >
> > > > The low-level API provides tools to build a feature,
> > > > but no specific feature.
> > > >
> > > > > > > For example, If we need to implement application facing" subsystems like bbdev,
> > > > > > > If we make all this driver interface, you can still implement the
> > > > > > > bbdev API as a driver without
> > > > > > > exposing HW specific details like how devices communicate to CPU, how
> > > > > > > memory is allocated etc
> > > > > > >  to "application".
> > > > > >
> > > > > > There are 2 things to understand here.
> > > > > >
> > > > > > First we want to allow the application using the GPU for needs which are
> > > > > > not exposed by any other DPDK API.
> > > > > >
> > > > > > Second, if we want to implement another DPDK API like bbdev,
> > > > > > then the GPU implementation would be exposed as a vdev in bbdev,
> > > > > > using the HW GPU device being a PCI in gpudev.
> > > > > > They are two different levels, got it?
> > > > >
> > > > > Exactly. So what is the point of exposing low-level driver API to
> > > > > "application",
> > > > > why not it is part of the internal driver API. My point is, why the
> > > > > application needs to worry
> > > > > about, How the CPU <-> Device communicated? CPU < -> Device memory
> > > > > visibility etc.
> > > >
> > > > There are two reasons.
> > > >
> > > > 1/ The application may want to use the GPU for some application-specific
> > > > needs which are not abstracted in DPDK API.
> > >
> > > Yes. Exactly, That's where my concern, If we take this path, What is
> > > the motivation to contribute to DPDK abstracted subsystem APIs which
> > > make sense for multiple vendors and every
> > > Similar stuff applicable for DPU,
> >
> > A feature-specific API is better of course, there is no lose of motivation.
> > But you cannot forbid applications to have their own features on GPU.
>
> it still can use it. We don't need DPDK APIs for that.
>
> >
> > > Otherway to put, if GPU is doing some ethdev offload, why not making
> > > as ethdev offload in ethdev spec so that
> > > another type of device can be used and make sense for application writters.
> >
> > If we do ethdev offload, yes we'll implement it.
> > And we'll do it on top of gpudev, which is the only way to share the CPU.
> >
> > > For example, In the future, If someone needs to add ML(Machine
> > > learning) subsystem and enable a proper subsystem
> > > interface that is good for DPDK. If this path is open, there is no
> > > motivation for contribution and the application
> > > will not have a standard interface doing the ML job across multiple vendors.
> >
> > Wrong. It does remove the motivation, it is a first step to build on top of it.
>
> IMO, No need to make driver API to the public to feature API.
>
> >
> > > That's is the only reason why saying it should not APPLICATION
> > > interface it can be DRIVER interface.
> > >
> > > >
> > > > 2/ This API may also be used by some feature implementation internally
> > > > in some DPDK libs or drivers.
> > > > We cannot skip the gpudev layer because this is what allows generic probing
> > > > of the HW, and gpudev allows to share the GPU with multiple features
> > > > implemented in different libs or drivers, thanks to the "child" concept.
> > >
> > > Again, why do applications need to know it? It is similar to `bus`
> > > kind of this where it sharing the physical resouces.
> >
> > No it's not a bus, it is a device that we need to share.
> >
> > > > > > > > > aka DPDK device class has a fixed personality and it has API to deal
> > > > > > > > > with abstracting specific application specific
> > > > > > > > > end user functionality like ethdev, cryptodev, eventdev irrespective
> > > > > > > > > of underlying bus/device properties.
> > > > > > > >
> > > > > > > > The goal of the lib is to allow anyone to invent any feature
> > > > > > > > which is not already available in DPDK.
> > > > > > > >
> > > > > > > > > Even similar semantics are required for DPU(Smart NIC)
> > > > > > > > > communitication. I am planning to
> > > > > > > > > send RFC in coming days to address the issue without the application
> > > > > > > > > knowing the Bus/HW/Driver details.
> > > > > > > >
> > > > > > > > gpudev is not exposing bus/hw/driver details.
> > > > > > > > I don't understand what you mean.
> > > > > > >
> > > > > > > See above.
> >
> > We are going into circles.
>
> Yes.
>
> > In short, Jerin wants to forbid the generic use of GPU in DPDK.
>
> See below.

Honestly I don’t see a real problem releasing the library at application level.
It doesn’t prevent to use it internally by other DPDK libraries/drivers if needed.

Applications can benefit of this library for a number of reasons:

  *   Enhance the interaction between GPU specific library and DPDK
  *   Hide GPU specific implementation details to the “final” user that wants to build a GPU + DPDK application
  *   Measure network throughput with common tools like testpmd using GPU memory

Please be aware that this is just a starting point.
I’m planning to expose a number of features (at memory and processing levels) that can be useful
to enhance the communication among GPU, CPU and NIC hiding the implementation
details within the library/driver.

>
> > He wants only feature-specific API.
>
> To re-reiterate, feature-specific "application" API. A device-specific
> bit can be
> driver API and accessible to the out-of-tree driver if needed.
>
> IMO, if we take this path, DPU, XPU, GPU, etc we need N different libraries to
> get the job done for a specific feature for the dataplane.
> Instead, Enabling public feature APIs will make the application
> portable and does not
> need to worry about which type of *PU it runs.
>

As I stated multiple times, let’s start with something simple that works and
then think about how to enhance the library/driver.
IMHO it doesn’t make sense to address all the use-cases now.
This is a completely new scenario we’re opening in the DPDK context, let’s start
from the basis.

>
> > It is like restricting the functions we can run on a CPU.
> >
> > And anyway we need this layer to share the GPU between multiple features.
>
> No disagreement there. Is that layer public application API or not is
> the question.
> it is like PCI device API calls over of the application and makes the
> application device specific.
>
> >
> > Techboard please vote.
>
> Yes.
>
> >
> >