running DPDK application on Azure

running DPDK application on Azure

[Yang Luan]

[-- Attachment #1: Type: text/plain, Size: 537 bytes --]

Hi,

We have an application using DPDK on AWS and would like to port it to
Azure. What would be recommended PMD to use? If I understand correctly, we
can either use the Netvsc PMD or the vdev_Netvsc PMD. It seems the Netvsc
PMD is newer.

An alternative is to use the mlx4 PMD by only attaching to the mlx NIC's
PCI address. As I understand it, the concern is the mlx nic may not receive
all the packets. We run a proprietary UDP based protocol on top of DPDK.
Are all UDP packets guaranteed to be received by the mlx NIC?

Thanks,
Yang

[-- Attachment #2: Type: text/html, Size: 645 bytes --]

Re: running DPDK application on Azure

[Stephen Hemminger]

On Tue, 12 Apr 2022 13:09:51 -0700
Yang Luan <luan.penny@gmail.com> wrote:

> Hi,
> 
> We have an application using DPDK on AWS and would like to port it to
> Azure. What would be recommended PMD to use? If I understand correctly, we
> can either use the Netvsc PMD or the vdev_Netvsc PMD. It seems the Netvsc
> PMD is newer.

Short answer:

Netvsc PMD is faster and can handle events better.
vdev_netvsc/failsafe/tap is slower but can emulate some types of rte_flow.

> 
> An alternative is to use the mlx4 PMD by only attaching to the mlx NIC's
> PCI address. As I understand it, the concern is the mlx nic may not receive
> all the packets. We run a proprietary UDP based protocol on top of DPDK.
> Are all UDP packets guaranteed to be received by the mlx NIC?

That won't work. the MLX device only sees established flows.

Re: running DPDK application on Azure

[Yang Luan]

[-- Attachment #1: Type: text/plain, Size: 1242 bytes --]

Thanks Stephen.

Is the Netvsc PMD selected by default or I'll need to specify it somewhere?

Since I'm running a proprietary UDP protocol, 3rd parties (e.g. Azure)
won't know how a flow is established. I'm curious how exactly Azure selects
which NIC to receive a given packet?

Yang

On Wed, Apr 13, 2022 at 8:58 AM Stephen Hemminger <
stephen@networkplumber.org> wrote:

> On Tue, 12 Apr 2022 13:09:51 -0700
> Yang Luan <luan.penny@gmail.com> wrote:
>
> > Hi,
> >
> > We have an application using DPDK on AWS and would like to port it to
> > Azure. What would be recommended PMD to use? If I understand correctly,
> we
> > can either use the Netvsc PMD or the vdev_Netvsc PMD. It seems the Netvsc
> > PMD is newer.
>
> Short answer:
>
> Netvsc PMD is faster and can handle events better.
> vdev_netvsc/failsafe/tap is slower but can emulate some types of rte_flow.
>
> >
> > An alternative is to use the mlx4 PMD by only attaching to the mlx NIC's
> > PCI address. As I understand it, the concern is the mlx nic may not
> receive
> > all the packets. We run a proprietary UDP based protocol on top of DPDK.
> > Are all UDP packets guaranteed to be received by the mlx NIC?
>
> That won't work. the MLX device only sees established flows.
>

[-- Attachment #2: Type: text/html, Size: 1787 bytes --]

Re: running DPDK application on Azure

[Stephen Hemminger]

On Wed, 13 Apr 2022 11:41:30 -0700
Yang Luan <luan.penny@gmail.com> wrote:

> Thanks Stephen.
> 
> Is the Netvsc PMD selected by default or I'll need to specify it somewhere?
> 
> Since I'm running a proprietary UDP protocol, 3rd parties (e.g. Azure)
> won't know how a flow is established. I'm curious how exactly Azure selects
> which NIC to receive a given packet?
> 
> Yang
> 
> On Wed, Apr 13, 2022 at 8:58 AM Stephen Hemminger <
> stephen@networkplumber.org> wrote:  
> 
> > On Tue, 12 Apr 2022 13:09:51 -0700
> > Yang Luan <luan.penny@gmail.com> wrote:
> >  
> > > Hi,
> > >
> > > We have an application using DPDK on AWS and would like to port it to
> > > Azure. What would be recommended PMD to use? If I understand correctly,  
> > we  
> > > can either use the Netvsc PMD or the vdev_Netvsc PMD. It seems the Netvsc
> > > PMD is newer.  
> >
> > Short answer:
> >
> > Netvsc PMD is faster and can handle events better.
> > vdev_netvsc/failsafe/tap is slower but can emulate some types of rte_flow.
> >  
> > >
> > > An alternative is to use the mlx4 PMD by only attaching to the mlx NIC's
> > > PCI address. As I understand it, the concern is the mlx nic may not  
> > receive  
> > > all the packets. We run a proprietary UDP based protocol on top of DPDK.
> > > Are all UDP packets guaranteed to be received by the mlx NIC?  
> >
> > That won't work. the MLX device only sees established flows.
> >  

The choice is made based on whether vmbus is unbound from kernel or not.
If vmbus is left bound to kernel, then hv_netvsc will look at all the network
devices, and those without any routes it will setup.  The setup makes the
TAP/failsafe sub devices.

If vmbus device is unbound from network device and instead bound to hv_uio_generic,
then the DPDK startup code will see it and should setup from there.

The documentation is there in many places.