On 26 Aug 2019, at 16:54, Ilya Maximets wrote:
'tx_q' array is allocated for each DPDK netdev. 'struct
dpdk_tx_queue'
is 8 bytes long, so 8 tx queues are sharing the same cache line in
case of 64B cacheline size. This causes 'false sharing' issue in
mutliqueue case because taking the spinlock implies write to memory
i.e. cache invalidation.
Signed-off-by: Ilya Maximets <[email protected]>
---
I didn't really test the performance impact yet, so testing is very
welcome. Relevant test case could be PVP with 8 queues in HW NIC and
8 queues in vhost-user inerface and 8 PMD threads.
Did a quick test on my setup, but I do not have enough cores to do 8 PMD
threads and 8 VM cores. So I did a 4 core PMD/VM test running PV (not
PVP as we do not care about receiving traffic from the VHOST) using 64
bytes packets and 100 L3 flows.
Without the patch, I get 17,408,585 pps and with the patch 17,573,089
pps.
The patch looks good to me…
Acked-by: Eelco Chaudron <[email protected]>
_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev
