[dpdk-dev] [PATCH] net/mlx5: remove excessive data prefetch

[dpdk-dev] [PATCH] net/mlx5: remove excessive data prefetch

[Yongseok Koh]

In Enhanced Multi-Packet Send (eMPW), entire packet data is prefetched to
LLC if it isn't inlined. Even though this helps reducing jitter when HW
fetches data by DMA, this can thresh the LLC with evicting precious data.
And if the size of queue is large and there are many queues, this might not
be effective. Also, if application runs on a remote node from the PCIe
link, it may not be helpful and can even cause bad results.

Signed-off-by: Yongseok Koh <yskoh@mellanox.com>
---
 drivers/net/mlx5/mlx5_rxtx.c | 8 ++------
 1 file changed, 2 insertions(+), 6 deletions(-)

diff --git a/drivers/net/mlx5/mlx5_rxtx.c b/drivers/net/mlx5/mlx5_rxtx.c
index 049f7e6c1..c2060b734 100644
--- a/drivers/net/mlx5/mlx5_rxtx.c
+++ b/drivers/net/mlx5/mlx5_rxtx.c
@@ -1320,7 +1320,6 @@ txq_burst_empw(struct mlx5_txq_data *txq, struct rte_mbuf **pkts,
 	do {
 		struct rte_mbuf *buf = *(pkts++);
 		uintptr_t addr;
-		unsigned int n;
 		unsigned int do_inline = 0; /* Whether inline is possible. */
 		uint32_t length;
 		uint8_t cs_flags;
@@ -1440,11 +1439,8 @@ txq_burst_empw(struct mlx5_txq_data *txq, struct rte_mbuf **pkts,
 					((uintptr_t)mpw.data.raw +
 					 inl_pad);
 			(*txq->elts)[elts_head++ & elts_m] = buf;
-			addr = rte_pktmbuf_mtod(buf, uintptr_t);
-			for (n = 0; n * RTE_CACHE_LINE_SIZE < length; n++)
-				rte_prefetch2((void *)(addr +
-						n * RTE_CACHE_LINE_SIZE));
-			addr = rte_cpu_to_be_64(addr);
+			addr = rte_cpu_to_be_64(rte_pktmbuf_mtod(buf,
+								 uintptr_t));
 			*dseg = (rte_v128u32_t) {
 				rte_cpu_to_be_32(length),
 				mlx5_tx_mb2mr(txq, buf),
-- 
2.11.0

Re: [dpdk-dev] [PATCH] net/mlx5: remove excessive data prefetch

[Adrien Mazarguil]

On Mon, Mar 12, 2018 at 10:05:45AM -0700, Yongseok Koh wrote:
> In Enhanced Multi-Packet Send (eMPW), entire packet data is prefetched to
> LLC if it isn't inlined. Even though this helps reducing jitter when HW
> fetches data by DMA, this can thresh the LLC with evicting precious data.
> And if the size of queue is large and there are many queues, this might not
> be effective. Also, if application runs on a remote node from the PCIe
> link, it may not be helpful and can even cause bad results.
> 
> Signed-off-by: Yongseok Koh <yskoh@mellanox.com>

Acked-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>

-- 
Adrien Mazarguil
6WIND

Re: [dpdk-dev] [PATCH] net/mlx5: remove excessive data prefetch

[Shahaf Shuler]

Tuesday, March 13, 2018 2:45 PM, Adrien Mazarguil:
> Subject: Re: [dpdk-dev] [PATCH] net/mlx5: remove excessive data prefetch
> 
> On Mon, Mar 12, 2018 at 10:05:45AM -0700, Yongseok Koh wrote:
> > In Enhanced Multi-Packet Send (eMPW), entire packet data is prefetched
> > to LLC if it isn't inlined. Even though this helps reducing jitter
> > when HW fetches data by DMA, this can thresh the LLC with evicting
> precious data.
> > And if the size of queue is large and there are many queues, this
> > might not be effective. Also, if application runs on a remote node
> > from the PCIe link, it may not be helpful and can even cause bad results.
> >
> > Signed-off-by: Yongseok Koh <yskoh@mellanox.com>
> 
> Acked-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>

Applied to next-net-mlx, thanks. 

> 
> --
> Adrien Mazarguil
> 6WIND