net/ice: enlarge ice RXQ rearm thresh to 64
Checks
Commit Message
We observe performance drop on ice AVX512 data path after stop and
start by using testpmd.
As CPU polling is faster in AVX512 path, L3 contested accesses is
intensified when rxrearm_start is a random value after testpmd
stop/start.
Enlarge ICE_RXQ_REARM_THRESH to 64 to ease the contested asscesses and
fix the performance drop issue.
Signed-off-by: Leyi Rong <leyi.rong@intel.com>
---
drivers/net/ice/ice_rxtx.h | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
Comments
> -----Original Message-----
> From: Rong, Leyi <leyi.rong@intel.com>
> Sent: Wednesday, January 6, 2021 1:36 PM
> To: Zhang, Qi Z <qi.z.zhang@intel.com>; Lu, Wenzhuo <wenzhuo.lu@intel.com>
> Cc: dev@dpdk.org; Rong, Leyi <leyi.rong@intel.com>
> Subject: [PATCH] net/ice: enlarge ice RXQ rearm thresh to 64
>
> We observe performance drop on ice AVX512 data path after stop and start by
> using testpmd.
>
> As CPU polling is faster in AVX512 path, L3 contested accesses is intensified
> when rxrearm_start is a random value after testpmd stop/start.
>
> Enlarge ICE_RXQ_REARM_THRESH to 64 to ease the contested asscesses and fix
> the performance drop issue.
>
> Signed-off-by: Leyi Rong <leyi.rong@intel.com>
Acked-by: Qi Zhang <qi.z.zhang@intel.com>
Applied to dpdk-next-net-intel after add cc:stable tag.
Thanks
Qi
@@ -31,7 +31,7 @@
#define ICE_VPMD_RX_BURST 32
#define ICE_VPMD_TX_BURST 32
-#define ICE_RXQ_REARM_THRESH 32
+#define ICE_RXQ_REARM_THRESH 64
#define ICE_MAX_RX_BURST ICE_RXQ_REARM_THRESH
#define ICE_TX_MAX_FREE_BUF_SZ 64
#define ICE_DESCS_PER_LOOP 4