[dpdk-dev] net/i40e: Significant performance degradation relative to DPDK 17.02 with TRex use-cases

[Alejandro Lucero]

We are seeing same performance drops but in our case is 16.11.3 compared
against 17.05.2 and 17.08.

That is when DPDK is used with SRIOV inside VMs, and the only change is the
DPDK version. Similar tests but using SRIOV in the host don't have such a
drop.

One change that could impact performance was added in 17.05, and it is
about how PMD private data is allocated. 16.11 and 17.02 use rte_zmalloc
which calls rte_zmalloc_socket using SOCKET_ID_ANY by default, but since
17.05 such private data is allocated using rte_zmalloc_socket with
dev->device.numa.node.

With some testing in the host, I can see a minor impact when a PMD is using
its private data allocated in another NUMA node, but this could have a
bigger impact inside VMs. I'm investigating this and will inform about any
progress.


On Fri, Nov 24, 2017 at 11:36 AM, Hanoch Haim (hhaim) <hhaim at cisco.com>
wrote:

> Hi All,
> Significant performance degradation observed with DPDK 17.11
> The Scenario is with TRex traffic generator  (https://github.com/cisco-
> system-traffic-generator/trex-core)
>
>
> 1.       Stateless mode, 64B packet stream (multi-core/single core)
>
> DPDK 17.02 - 37-39MPPS/core
> DPDK 17.11 - 33.5MPPS/core
>
> ~12% degradation
>
>
> 2.       Stateless mode, Field Engine  (Multi-core)
>
>
> DPDK 17.02 - 14MPPS/core
> DPDK 17.11 -11.8MPPS/core
>
> ~15% degradation
>
>
> 3.       Only DPDK was replaced
>
> 4.       Is this expected?
>
> The full report can be found here:
> https://trex-tgn.cisco.com/trex/doc/trex_analytics.html
>
>
>
> *         Mellanox CX-5 (mlx5) shows significant improvement
>
> *         Ixgbe shows improvement
>
> Thanks,
> Hanoh
>
>