[dpdk-dev] Queue Management Support in DPDK

[dpdk-dev] Queue Management Support in DPDK

[Archit Pandey]

Hello everyone,

We have been using DPDK's QoS framework over the last year and found
that rte_sched and the provided qos_sched app work great for QoS.

However, when we ventured into trying to add CoDel (to replace RED) as
a dropper to the framework, we faced several challenges due to how
tightly rte_sched and rte_red were coupled together. As we had no
success with rte_sched, we would like to propose a new framework for
queue management in DPDK.

Goals we have in mind for the framework:
- Act as an abstraction for queue management algorithms (AQMs) such as
CoDel, PiE and RED.
- Make it easy for new algorithms to be added.

We’d appreciate feedback on whether such a framework would be welcomed
in the community, or what else could be done for adding queue
management support.

Sincerely,
Archit Pandey.

Re: [dpdk-dev] Queue Management Support in DPDK

[Stephen Hemminger]

On Thu, 30 Jul 2020 00:07:29 +0530
Archit Pandey <architpandeynitk@gmail.com> wrote:

> Hello everyone,
> 
> We have been using DPDK's QoS framework over the last year and found
> that rte_sched and the provided qos_sched app work great for QoS.
> 
> However, when we ventured into trying to add CoDel (to replace RED) as
> a dropper to the framework, we faced several challenges due to how
> tightly rte_sched and rte_red were coupled together. As we had no
> success with rte_sched, we would like to propose a new framework for
> queue management in DPDK.
> 
> Goals we have in mind for the framework:
> - Act as an abstraction for queue management algorithms (AQMs) such as
> CoDel, PiE and RED.
> - Make it easy for new algorithms to be added.
> 
> We’d appreciate feedback on whether such a framework would be welcomed
> in the community, or what else could be done for adding queue
> management support.
> 
> Sincerely,
> Archit Pandey.

rte_sched is not a generic AQM mechanism. You will have to write a new
replacement for rte_sched if you want something else.

I would recommend starting with Cake. It is latest and most complete
and the developers are active and friendly.

Re: [dpdk-dev] Queue Management Support in DPDK

[Archit Pandey]

Hi Stephen,

Thanks for the reply.

On Thu, Jul 30, 2020 at 12:37 AM Stephen Hemminger
<stephen@networkplumber.org> wrote:
>
> rte_sched is not a generic AQM mechanism. You will have to write a new
> replacement for rte_sched if you want something else.

Yes, we did realize that rte_sched is not suitable for AQM algorithms.
We have implemented a bare-bones AQM mechanism that is independent of rte_sched.
Using this implementation we have successfully added and tested PIE and CoDel.

> I would recommend starting with Cake. It is latest and most complete
> and the developers are active and friendly.

CAKE can be added using what we have built, but before we do that we wanted
to know if such a feature would be welcomed by the community. Subsequently,
we will submit an RFC patch series to the dev mailing list.

Thanks,
Archit Pandey.

On Thu, Jul 30, 2020 at 12:37 AM Stephen Hemminger
<stephen@networkplumber.org> wrote:
>
> On Thu, 30 Jul 2020 00:07:29 +0530
> Archit Pandey <architpandeynitk@gmail.com> wrote:
>
> > Hello everyone,
> >
> > We have been using DPDK's QoS framework over the last year and found
> > that rte_sched and the provided qos_sched app work great for QoS.
> >
> > However, when we ventured into trying to add CoDel (to replace RED) as
> > a dropper to the framework, we faced several challenges due to how
> > tightly rte_sched and rte_red were coupled together. As we had no
> > success with rte_sched, we would like to propose a new framework for
> > queue management in DPDK.
> >
> > Goals we have in mind for the framework:
> > - Act as an abstraction for queue management algorithms (AQMs) such as
> > CoDel, PiE and RED.
> > - Make it easy for new algorithms to be added.
> >
> > We’d appreciate feedback on whether such a framework would be welcomed
> > in the community, or what else could be done for adding queue
> > management support.
> >
> > Sincerely,
> > Archit Pandey.
>
> rte_sched is not a generic AQM mechanism. You will have to write a new
> replacement for rte_sched if you want something else.
>
> I would recommend starting with Cake. It is latest and most complete
> and the developers are active and friendly.



-- 
Archit Pandey
Senior Year B.Tech.
Department of Computer Science and Engineering
National Institute of Technology Karnataka
Surathkal, India

On Thu, Jul 30, 2020 at 12:37 AM Stephen Hemminger
<stephen@networkplumber.org> wrote:
>
> On Thu, 30 Jul 2020 00:07:29 +0530
> Archit Pandey <architpandeynitk@gmail.com> wrote:
>
> > Hello everyone,
> >
> > We have been using DPDK's QoS framework over the last year and found
> > that rte_sched and the provided qos_sched app work great for QoS.
> >
> > However, when we ventured into trying to add CoDel (to replace RED) as
> > a dropper to the framework, we faced several challenges due to how
> > tightly rte_sched and rte_red were coupled together. As we had no
> > success with rte_sched, we would like to propose a new framework for
> > queue management in DPDK.
> >
> > Goals we have in mind for the framework:
> > - Act as an abstraction for queue management algorithms (AQMs) such as
> > CoDel, PiE and RED.
> > - Make it easy for new algorithms to be added.
> >
> > We’d appreciate feedback on whether such a framework would be welcomed
> > in the community, or what else could be done for adding queue
> > management support.
> >
> > Sincerely,
> > Archit Pandey.
>
> rte_sched is not a generic AQM mechanism. You will have to write a new
> replacement for rte_sched if you want something else.
>
> I would recommend starting with Cake. It is latest and most complete
> and the developers are active and friendly.



-- 
Archit Pandey
Senior Year B.Tech.
Department of Computer Science and Engineering
National Institute of Technology Karnataka
Surathkal, India

Re: [dpdk-dev] Queue Management Support in DPDK

[Jon Loeliger]

On Sun, Aug 2, 2020 at 4:43 AM Archit Pandey <architpandeynitk@gmail.com>
wrote:

> Hi Stephen,
>
> Thanks for the reply.
>
> On Thu, Jul 30, 2020 at 12:37 AM Stephen Hemminger
> <stephen@networkplumber.org> wrote:
> >
> > rte_sched is not a generic AQM mechanism. You will have to write a new
> > replacement for rte_sched if you want something else.
>
> Yes, we did realize that rte_sched is not suitable for AQM algorithms.
> We have implemented a bare-bones AQM mechanism that is independent of
> rte_sched.
> Using this implementation we have successfully added and tested PIE and
> CoDel.
>
> > I would recommend starting with Cake. It is latest and most complete
> > and the developers are active and friendly.
>
> CAKE can be added using what we have built, but before we do that we wanted
> to know if such a feature would be welcomed by the community.


In a word, yes.


> Subsequently, we will submit an RFC patch series to the dev mailing list.
>
> Thanks,
> Archit Pandey.
>

HTH,
jdl