NVIDIA roadmap for 25.03

NVIDIA roadmap for 25.03

[Maayan Kashani]

[-- Attachment #1: Type: text/plain, Size: 2437 bytes --]

Please find below NVIDIA roadmap for 25.03 release:



EAL change

=============

Improve PCI device string comparisons by using bus-specific parsing functions.

By comparing the parsed PCI address instances instead of the string representations,

the application can ensure consistent handling of PCI device numbers, regardless of the format used.



rte_ethdev new APIs  

=================  

Hairpin Queue Statistics query​  

Track detailed statistics for hairpin operations at the queue level, including counters on receive misses, transmission errors, and queue drops.​  

This enhanced monitoring capability enables improved debugging, telemetry collection, and performance tracking. 

Hardware support and capabilities:

  *   ConnectX-6 Dx: Statistics available at port level and Physical Function (PF).

  *   ConnectX-7, BlueField-3 and newer: Extended support for per-queue statistics and Virtual Function (VF) monitoring.



Net/mlx5 PMD updates  

=====================  

  1.  Forwarding Database (FDB) domain tables​ Optimization 

Network flows in FDB often lack directional indicators,

resulting in redundant flow entries created in both receive (Rx) and transmit (Tx) domains.

This has been changed with the introduction of a unified FDB approach which addresses the inefficiency.

Therefore, the unified FDB reduces the duplication of Tx / Rx tables in the FDB domain ,

optimize the performance and reduce memory usage to allow room for more offloads (increase total scale).

  1.  Optimization of device probing time

Probing time with hundreds of REPs is too long, this optimization is cashing the probing data and using a new kernel API.

  1.  Support eCPRI matching in template API.

Support eCPRI based on Flex Parser

  1.  Support queue and RSS actions from FDB

Save the hops in Rx direction for performance, bypass NIC Rx domain

  1.  Memory pool debuggability

Introducing live tracing at multiple log levels for operations within the ipool, providing detailed visibility into its internal structures and processes.

This feature also includes checks for double malloc/free operations to enhance reliability and debuggability.



Test apps updates    

=================   

Support the changes in rte_flow listed above in testpmd.   




Regards,
Maayan Kashani


[-- Attachment #2: Type: text/html, Size: 16334 bytes --]

NVIDIA roadmap for 25.07

[Maayan Kashani]

[-- Attachment #1: Type: text/plain, Size: 1906 bytes --]

Please find below NVIDIA roadmap for 25.07 release:

EAL change
=============
Improve PCI device string comparisons by using bus-specific parsing functions.
Ensure consistent handling of PCI device numbers, regardless of the format used by comparing the parsed PCI
address instances instead of the string representations.

Net/mlx5 PMD updates  
=====================  

  1.  Tx datapath optimization
Optimize Tx datapath performance by changing Tx inner buffers allocation and alignment in memory.

  1.  Memory statistics visibility for debugging
                Adding new xstats counters for debugging memory usage and allocations.

  1.  Support for legacy rte_flow (non-template) on top of the async hardware mechanism (HWS)
has been added as a beta feature, serving as a preview for future NICs (e.g., ConnectX-9+),
where legacy rte_flow will no longer be supported,
and will be exchanged with the [already available] template async interface.
While this mode is compatible with older NICs (prior to ConnectX-9),
it is not recommended since existing legacy mode delivers better performance.
Developers are strongly encouraged to transition to the template API for both
superior performance and access to an enhanced feature set.

  1.  Handle SF/VF representor removal in DPDK (supported with ConnectX-6DX and above)
Listen to the kernel events and remove needed SFs/VFs representors.

  1.  Cross Numa support
Enable fallback to any socket to support cross NUMA memory allocations during initialization.

  1.  Support matching on IPv6 frag extension header with rte_flow template API.

  1.  Support RTE_FLOW_ACTION_TYPE_COUNT on group 0 (root) in rte_flow template API.

  1.  Support IP-in-IP matching in rte_flow template API.

Doc
=======
Updated mlx5 documentation with feature description and usage examples.


Regards,
Maayan Kashani

[-- Attachment #2: Type: text/html, Size: 13776 bytes --]