[dpdk-dev] DPDK 20.05 Mellanox Roadmap

[dpdk-dev] DPDK 20.05 Mellanox Roadmap

[Gal Cohen]

Below is Mellanox's roadmap for DPDK20.05:



Reduce memory consumption in mlx5 PMD -

[1] Reduce flow memory (entry size) footprint/consumption.

[2] Remove flow rules caching.

Change the mlx5 PMD implementation of rte_eth_dev_stop()/rte_eth_dev_start() to stop caching flow rules (freeing resources for the created flows).

Benefits: Scale and performance improvement.



HW offload for TTL matching

[3] Offload TTL matching from routing applications to the NIC.

Usage through rte_flow API ; Implementation in mlx5 PMD.

Benefits: Simplifying the routing application complexity as well as reducing host CPU usage through offloading the TTL logic to the NIC.



Flow aging - Introducing new AGE action in the rte_flow API

[4]  Add a new action to allow the application (client) to define an age threshold (seconds) on which it expects to get notification from the mlx5 PMD.

The PMD will implement the flow aging monitoring aging through the use of the rte_flow API, instead of the application.

Additional background can be found in here: https://patchwork.dpdk.org/patch/53701/.



Add support for additional features in vDPA

            [5] Support Large-Send-Offload (LSO).

            [6] Improve debuggability through vDPA device counters and error reporting.

            [7] Rate limiting - allow application definition of the maximal number of packets; providing means to apply different policies/QoS for different applications / tenants.

            [8] ConnectX-6 Dx - introduce HW Doorbell instead of current SW Doorbell.



Performance improvement for jumbo-frame size

            [9] Add support for large MTU size while using MPRQ to provide full line rate performance with any packet size.



Hairpin with jumbo-frames

 [10] Add support for jumbo-frame packets In addition to the existing hairpin offloading functionality.



Enable GTP encap/decap

[11] Add encap/decap to GTP using the rte_flow raw_encap/raw_decap APIs.



Adding test application

           [12] Introduce a new (standalone) test application for measuring flow insertion rate and traffic performance.

           The application defines flows with rte_flow API ; Measure the flow insertion rate and traffic performance.

Re: [dpdk-dev] DPDK 20.05 Mellanox Roadmap

[Thomas Monjalon]

Summarized for the web:
	http://git.dpdk.org/tools/dpdk-web/commit/?id=7edb98830


16/03/2020 09:21, Gal Cohen:
> Below is Mellanox's roadmap for DPDK20.05:
> 
> 
> 
> Reduce memory consumption in mlx5 PMD -
> 
> [1] Reduce flow memory (entry size) footprint/consumption.
> 
> [2] Remove flow rules caching.
> 
> Change the mlx5 PMD implementation of rte_eth_dev_stop()/rte_eth_dev_start() to stop caching flow rules (freeing resources for the created flows).
> 
> Benefits: Scale and performance improvement.
> 
> 
> 
> HW offload for TTL matching
> 
> [3] Offload TTL matching from routing applications to the NIC.
> 
> Usage through rte_flow API ; Implementation in mlx5 PMD.
> 
> Benefits: Simplifying the routing application complexity as well as reducing host CPU usage through offloading the TTL logic to the NIC.
> 
> 
> 
> Flow aging - Introducing new AGE action in the rte_flow API
> 
> [4]  Add a new action to allow the application (client) to define an age threshold (seconds) on which it expects to get notification from the mlx5 PMD.
> 
> The PMD will implement the flow aging monitoring aging through the use of the rte_flow API, instead of the application.
> 
> Additional background can be found in here: https://patchwork.dpdk.org/patch/53701/.
> 
> 
> 
> Add support for additional features in vDPA
> 
>             [5] Support Large-Send-Offload (LSO).
> 
>             [6] Improve debuggability through vDPA device counters and error reporting.
> 
>             [7] Rate limiting - allow application definition of the maximal number of packets; providing means to apply different policies/QoS for different applications / tenants.
> 
>             [8] ConnectX-6 Dx - introduce HW Doorbell instead of current SW Doorbell.
> 
> 
> 
> Performance improvement for jumbo-frame size
> 
>             [9] Add support for large MTU size while using MPRQ to provide full line rate performance with any packet size.
> 
> 
> 
> Hairpin with jumbo-frames
> 
>  [10] Add support for jumbo-frame packets In addition to the existing hairpin offloading functionality.
> 
> 
> 
> Enable GTP encap/decap
> 
> [11] Add encap/decap to GTP using the rte_flow raw_encap/raw_decap APIs.
> 
> 
> 
> Adding test application
> 
>            [12] Introduce a new (standalone) test application for measuring flow insertion rate and traffic performance.
> 
>            The application defines flows with rte_flow API ; Measure the flow insertion rate and traffic performance.
>