[Public]
Hi Morten,
We have tested the effect of the patch using func-latency and PPs via testpmd.
Please find our observations below
- DPDK tag: 25.07-rc1
- compiler: gcc 14.2
- platform: AMD EPYC 8534P 64core 2.3GHz
- app cmd:
-- One port: ` sudo build/app/dpdk-testpmd -l 15,16 --vdev=net_null1 --no-pci
-- --nb-cores=1 --nb-ports=1 --txq=1 --rxq=1 --txd=2048 --rxd=2048 -a
--forward-mode=io --stats-period=1`
-- Two port: ` sudo build/app/dpdk-testpmd -l 15,16,17 --vdev=net_null1
--vdev=net_null2 --no-pci -- --nb-cores=2 --nb-ports=2 --txq=1 --rxq=1
--txd=2048 --rxd=2048 -a --forward-mode=io --stats-period=1`
Result 1 port:
- Before patch: TX MPPs 117.61, RX-PPs 117.67, Func-latency TX: 1918ns,
Func-latency free-bulk: 2667ns
- After patch: TX MPPs 117.55, RX-PPs 117.54, Func-latency TX: 1921ns,
Func-latency free-bulk: 2660ns
Result 2 port:
- Before patch: TX MPPs 117.61, RX-PPs 117.67, Func-latency TX: 1942ns,
Func-latency free-bulk: 2557ns
- After patch: TX MPPs 117.54, RX-PPs 117.54, Func-latency TX: 1946ns,
Func-latency free-bulk: 2740ns
Perf Top: diff before vs after shows 13.84% vs 13.79%
Reviewed-by: Thiyagarjan P <thiyagaraja...@amd.com>
Tested-by: Vipin Varghese <vipin.vargh...@amd.com>
Clarification request: `with fast-mbuf-free on single port we see free-bulk
reduction by -7ns. But null_tx increase by +3ns. TX PPs reduction by 0.07 Mpps.
Is this anomaly of null_net PMD?`
> >
> > On Tue, 24 Jun 2025 18:14:16 +0000
> > Morten Brørup <m...@smartsharesystems.com> wrote:
> >
> > > Added fast mbuf release, re-using the existing mbuf pool pointer in
> > > the queue structure.
> > >
> > > Signed-off-by: Morten Brørup <m...@smartsharesystems.com>
> >
> > Makes sense.
> >
> > > ---
> > > drivers/net/null/rte_eth_null.c | 30 +++++++++++++++++++++++++++---
> > > 1 file changed, 27 insertions(+), 3 deletions(-)
> > >
> > > diff --git a/drivers/net/null/rte_eth_null.c
> > b/drivers/net/null/rte_eth_null.c
> > > index 8a9b74a03b..12c0d8d1ff 100644
> > > --- a/drivers/net/null/rte_eth_null.c
> > > +++ b/drivers/net/null/rte_eth_null.c
> > > @@ -34,6 +34,17 @@ struct pmd_internals; struct null_queue {
> > > struct pmd_internals *internals;
> > >
> > > + /**
> > > + * For RX queue:
> > > + * Mempool to allocate mbufs from.
> > > + *
> > > + * For TX queue:
> > > + * Mempool to free mbufs to, if fast release of mbufs is enabled.
> > > + * UINTPTR_MAX if the mempool for fast release of mbufs has not
> > yet been detected.
> > > + * NULL if fast release of mbufs is not enabled.
> > > + *
> > > + * @see RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE
> > > + */
> > > struct rte_mempool *mb_pool;
> >
> > Do all drivers to it this way?
>
> No, I think most drivers have separate structures for rx and tx queues. This
> driver
> doesn't so I'm reusing the existing mempool pointer.
> Also, they don't cache the mempool pointer, but look at mbuf[0].pool at every
> burst;
> so their tx queue structure doesn't have a mempool pointer field.
> And they check an offload flag (either the bit in the raw offload field, or a
> shadow
> variable for the relevant offload flag), instead of checking the mempool
> pointer.
>
> Other drivers can be improved, and I have submitted an optimization patch for
> the
> i40e driver with some of the things I do in this patch:
> https://inbox.dpdk.org/dev/20250624061238.89259-1-
> m...@smartsharesystems.com/
>
> > Is it documented in ethdev?
>
> The RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE flag is documented.
> How to implement it in the drivers is not.
>
> -Morten
