[PATCH 00/11] net/mlx5: flow insertion performance improvements
Ori Kam
orika at nvidia.com
Thu Feb 29 09:52:34 CET 2024
Hi Dariusz,
> -----Original Message-----
> From: Dariusz Sosnowski <dsosnowski at nvidia.com>
> Sent: Wednesday, February 28, 2024 7:01 PM
>
> Goal of this patchset is to improve the throughput of flow insertion
> and deletion in mlx5 PMD when HW Steering flow engine is used.
>
> - Patch 1 - Use preallocated per-queue, per-actions template buffer
> for storing translated flow actions, instead of allocating and
> filling it on demand, on each flow operation.
> - Patches 2-4 - Make resource index allocation optional. This allocation
> will be skipped when it is not required by the created template table.
> - Patches 5-7 - Reduce memory footprint of the internal flow queue.
> - Patch 8 - Remove indirection between flow job and flow itself,
> by using flow as an operation container.
> - Patches 9-10 - Reduce memory footpring of flow struct by moving
> rarely used flow fields outside of the main flow struct.
> These fields will accesses only when needed.
> Also remove unneeded `zmalloc` usage.
> - Patch 11 - Remove unneeded device status check in flow create.
>
> In general all of these changes result in the following improvements
> (all numbers are averaged Kflows/sec):
>
> | | Insertion) | +% | Deletion | +% |
> |--------------|:----------:|:------:|:--------:|:-----:|
> | baseline | 6338.7 | | 9739.6 | |
> | improvements | 6978.8 | +10.1% | 10432.4 | +7.1% |
>
> The basic benchmark was run on ConnectX-6 Dx (22.40.1000),
> on the system with Intel Xeon Platinum 8380 CPU.
>
> Bing Zhao (2):
> net/mlx5: skip the unneeded resource index allocation
> net/mlx5: remove unneeded device status checking
>
> Dariusz Sosnowski (7):
> net/mlx5: allocate local DR rule action buffers
> net/mlx5: remove action params from job
> net/mlx5: remove flow pattern from job
> net/mlx5: remove updated flow from job
> net/mlx5: use flow as operation container
> net/mlx5: move rarely used flow fields outside
> net/mlx5: reuse flow fields
>
> Erez Shitrit (2):
> net/mlx5/hws: add check for matcher rule update support
> net/mlx5/hws: add check if matcher contains complex rules
>
> drivers/net/mlx5/hws/mlx5dr.h | 16 +
> drivers/net/mlx5/hws/mlx5dr_action.c | 6 +
> drivers/net/mlx5/hws/mlx5dr_action.h | 2 +
> drivers/net/mlx5/hws/mlx5dr_matcher.c | 29 +
> drivers/net/mlx5/mlx5.h | 29 +-
> drivers/net/mlx5/mlx5_flow.h | 128 ++++-
> drivers/net/mlx5/mlx5_flow_hw.c | 794 ++++++++++++++++----------
> 7 files changed, 666 insertions(+), 338 deletions(-)
>
> --
> 2.39.2
Series-acked-by: Ori Kam <orika at nvidia.com>
Best,
Ori
More information about the dev
mailing list