[ofa-general] [PATCH 10/11]: IB/ipoib modify cq params
Eli Cohen
eli at mellanox.co.il
Tue Sep 25 07:31:18 PDT 2007
> I think we have to carefully think if/how does this feature goes hand in
> hand with NAPI. Since NAPI is not optional, with this feature the
> network stack tries to do its best to reduce interrupts with the NAPI
> logic, and on top of that the HW is instructed to apply this logic
> before issuing an interrupt.
My experience shows that interrupt moderation helps make better us of
NAPI. Without it, chances are higher that the device generates an
interrupt request, a fast CPU polls the CQE and returns after calling
netif_rx_complete (which will cause the next CQE to generate another
interrupt request). When using moderation it is more probable that the
CQ will already contain more CQEs that will cause NAPI to contribute
even more to coalescing.
>
> Does the need here suggests that NAPI can be improved? if yes how? maybe
> for some infiniband devices interrupt moderation for itself would be
> better so NAPI should be disabled?
I think you can diminish the effect of NAPI by setting the weight
parameter to a lower value.
Not all devices support interrupt moderation.
>
> To suggest this for merge, I think you would need to share the list with
> the IPoIB results you had with NAPI vs with NAPI AND interrupt moderation.
I don't have orderly records of the effect but I can verify that using
moderation does improve performance. Anyone can experiment with it.
You can use ethtool to set different values.
>
> Or.
>
More information about the general
mailing list