On 4/6/2018 2:20 PM, Ferruh Yigit wrote: > On 4/4/2018 2:42 PM, Matej Vido wrote: >> Previously the queues were the part of private data structure of the >> Ethernet device. >> Now the queues are allocated at setup thus numa-aware allocation is >> possible. > > Hi Matej, > > Yes by default [rt]x_queues are allocated via rte_zmalloc, which uses > SOCKET_ID_ANY. > > And in burst functions, we do: > nb_rx = (*dev->rx_pkt_burst)(dev->data->rx_queues[queue_id], rx_pkts, > nb_pkts); > > So there is an access to rx_queues in each rte_eth_rx_burst() call. > > I wonder if you observe any performance difference with this update? > And what about moving to the ethdev layer instead of keeping local to the PMD?
Forget about it, I thought you are allocating data->[rt]x_queues in specific socket, but this just allocating queues in specific socket, which is OK. Still I would like to hear comments if allocating data->[rt]x_queues arrays in specific socket helps for performance. > >> >> Signed-off-by: Matej Vido <v...@cesnet.cz> > > <...> >