On Mon, Apr 03, 2017 at 10:47:20AM +0000, Ananyev, Konstantin wrote: > > > > -----Original Message----- > > From: dev [mailto:dev-boun...@dpdk.org] On Behalf Of Pei, Yulong > > Sent: Saturday, April 1, 2017 3:02 AM > > To: Vladyslav Buslov <vladyslav.bus...@harmonicinc.com>; Zhang, Helin > > <helin.zh...@intel.com>; Wu, Jingjing <jingjing...@intel.com>; > > Yigit, Ferruh <ferruh.yi...@intel.com> > > Cc: dev@dpdk.org > > Subject: Re: [dpdk-dev] [PATCH] net/i40e: add packet prefetch > > > > Hi All > > > > In Non-vector mode, without this patch, single core performance can reach > > 37.576Mpps with 64Byte packet, > > But after applied this patch , single core performance downgrade to > > 34.343Mpps with 64Byte packet. > > > > Best Regards > > Yulong Pei > > > > -----Original Message----- > > From: dev [mailto:dev-boun...@dpdk.org] On Behalf Of Vladyslav Buslov > > Sent: Wednesday, March 1, 2017 6:57 PM > > To: Zhang, Helin <helin.zh...@intel.com>; Wu, Jingjing > > <jingjing...@intel.com>; Yigit, Ferruh <ferruh.yi...@intel.com> > > Cc: dev@dpdk.org > > Subject: [dpdk-dev] [PATCH] net/i40e: add packet prefetch > > > > Prefetch both cache lines of mbuf and first cache line of payload if > > CONFIG_RTE_PMD_PACKET_PREFETCH is set. > > > > Signed-off-by: Vladyslav Buslov <vladyslav.bus...@harmonicinc.com> > > --- > > drivers/net/i40e/i40e_rxtx.c | 20 ++++++++++++++++---- > > 1 file changed, 16 insertions(+), 4 deletions(-) > > > > diff --git a/drivers/net/i40e/i40e_rxtx.c b/drivers/net/i40e/i40e_rxtx.c > > index 48429cc..2b4e5c9 100644 > > --- a/drivers/net/i40e/i40e_rxtx.c > > +++ b/drivers/net/i40e/i40e_rxtx.c > > @@ -100,6 +100,12 @@ > > #define I40E_TX_OFFLOAD_NOTSUP_MASK \ > > (PKT_TX_OFFLOAD_MASK ^ I40E_TX_OFFLOAD_MASK) > > > > +#ifdef RTE_PMD_PACKET_PREFETCH > > +#define rte_packet_prefetch(p) rte_prefetch0(p) > > +#else > > +#define rte_packet_prefetch(p) do {} while (0) > > +#endif > > + > > static uint16_t i40e_xmit_pkts_simple(void *tx_queue, > > struct rte_mbuf **tx_pkts, > > uint16_t nb_pkts); > > @@ -495,6 +501,9 @@ i40e_rx_scan_hw_ring(struct i40e_rx_queue *rxq) > > /* Translate descriptor info to mbuf parameters */ > > for (j = 0; j < nb_dd; j++) { > > mb = rxep[j].mbuf; > > + rte_packet_prefetch( > > + RTE_PTR_ADD(mb->buf_addr, > > + RTE_PKTMBUF_HEADROOM)); > > qword1 = rte_le_to_cpu_64(\ > > rxdp[j].wb.qword1.status_error_len); > > pkt_len = ((qword1 & I40E_RXD_QW1_LENGTH_PBUF_MASK) >> > > @@ -578,9 +587,11 @@ > > i40e_rx_alloc_bufs(struct i40e_rx_queue *rxq) > > > > rxdp = &rxq->rx_ring[alloc_idx]; > > for (i = 0; i < rxq->rx_free_thresh; i++) { > > - if (likely(i < (rxq->rx_free_thresh - 1))) > > + if (likely(i < (rxq->rx_free_thresh - 1))) { > > /* Prefetch next mbuf */ > > - rte_prefetch0(rxep[i + 1].mbuf); > > + rte_packet_prefetch(rxep[i + 1].mbuf->cacheline0); > > + rte_packet_prefetch(rxep[i + 1].mbuf->cacheline1); > > As I can see the line aove is the only real difference in that patch. > If that so, might be worth to re-run perf tests witout that line? > Konstantin > The prefetch for the packet buf_addr+headroom above also looks new. Are both needed to get the performance boost you see? We should also investigate if the same effect can be got using a runtime option, rather than a compile-time setting. That would give us the best of both worlds.
/Bruce
