Sagi,
Thanks for you patches, Max and me ran your patches adding completion
signaling for both send (command capsule) and local invalidation request.
The results were significant reduction of IOPs due to interrupts and
completion processing, see below our results for READ/WRITE IOPs (512B,
64 Jobs):
(1) No send signal and no invalidation signal w/ moderation 7.4M / 11.9M
(2) No send signal and no invalidation signal w/o moderation 7.1M / 8.7M
(3) No send signal with invalidation signal w/o moderation 5.2M / 8.3M
(4) Send signal with invalidation signal w/o moderation 3.4M / 4.6M
(5) Send signal with invalidation signal w/ moderation 6.2M / 11.8M
I understand that (1), (2) and (3) are invalid, the performance
reduction induce by mode (4) need to motivate us to enable cqe/event
moderation as soon as possible.
On 10/31/2017 1:10 PM, Sagi Grimberg wrote:
Thanks Sagi.
We'll run these patches in our lab + our patches for the performance
improvments that we sent few days ago and update the results.
Thanks,
I guess I'll need to run it with my cq moderation patches to get less
penalty in performance.
Fine, but just a heads up, any cq moderation that is not adaptive is
not something
that we'd like to pursue. I have a generic adaptive moderation library
in the
works, so any ideas should be integrated into that (or something
equivalent).
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
