On 2023-10-06 14:19, Bruce Richardson wrote:
On Fri, Oct 06, 2023 at 02:15:00PM +0200, Mattias Rönnblom wrote:
On 2023-10-05 13:51, Bruce Richardson wrote:
The event structure in DPDK is 16-bytes in size, and events are
regularly passed as parameters directly rather than being passed as
pointers.
When are events passed by-value, rather than by-reference? There are no such
examples in the public eventdev API.
To help compiler optimize correctly, we can explicitly request
16-byte alignment for events, which means that we should be able
to do aligned vector loads/stores (e.g. with SSE or Neon) when working
with those events.
That change is both helping and sabotaging the optimizer's work. Now every
stack allocation needs to be 2-byte aligned - in DPDK code, and in the
application. >>
The effect this change has on an eventdev app using DSW is a ~3 cycle/event
performance degradation on an AMD Zen 3 system, and a ~4 cycle/event
performance degradation on a Skylake-generation Intel CPU.
Thanks for checking - this is the sort of feedback needed alright. In SW
eventdev we copy events around alot without using pointers, so I felt that
alignment would be helpful to avoid issues with events spanning cachelines.
However, since it has negative impacts, I'm quite happy to drop the idea,
and keep things as they are. I'll mark the change as rejected in patchwork.
I think this was an excellent idea, it just didn't happen to have the
desired effect. At least not in the particular cases where I evaluated it.
Given the complexity of compilers and CPUs, it's almost impossible to
predict the outcome, performance-wise, of a particular change.
