Hi Ying,
 

It seems all the TxQs are sharing the same interrupt for Tx
completions. Will it be better to have separate interrupt per
num_tx_rings_p_up (8) queues? E.g. for a 16 core system, with 16 * 8
Tx queues, to have 16 interrupts for Tx completions of those 128 Tx
queues?

Actually not all TxQs share same interrupt vector.
In commit 76532d0c we assigned an interrupt vector for each TX ring.
When the number of Queues is higher than number of interrupt vectors, there are queues that share interrupts
And actually reaching the assignment you specified.

Also I'm looking at mlx4_en_select_queue(), it is using
__skb_tx_hash(). Use something to achieve XPS may bring better
performances.

We are considering this change.