On 2019-12-06 23:22, Venky Venkatesh wrote:
To my understanding, per eventdev API, events are considered in flight
between NEW to RELEASE (implicit/explicit).
DSW considers events to be in flight between time of enqueue (on the
source port), to the time of release (on the destination port). This is
regardless if they are FORWARD or NEW-type events.
The SW event device only counts NEW events toward the max number of
in-flights events limitation. It will however release the atomic context
and return the credits at step 4 below. Since there's no centralized
point which can backpressure FORWARD events (which I think is how SW
avoids being overwhelmed by FORWARDed events), DSW needs to ensure that
the max_num_events is never exceeded.
Now consider an event (event-1) going thru the following stages:
1. NEW from core-3
2. dequeued by core-1
3. FORWARD
4. core-1 does a next dequeue
The event is RELEASEd already at this point.
5. dequeued by core-2
6. RELEASE by core-2/implicit release on next dequeue by core-2
The way I understand DSW implementation this event would use credit at step
1 AND step 3 while releasing in step2 -- right now credit usage is for
non_release (i.e NEW and FORWARD). So if between step-2 and step-3 another
core puts in a NEW of event-2 that could utilize all the credits of the
system and could thus fail step-3 of event-1.
The NEW events producer ports' new_event_threshold should be
significantly lower than the maximum numbers of in-flight events, and
thus leave credits for FORWARDed events.
This to my knowledge is not conformant with eventdev. One way to address
this is to track the credits for that which are currently in core and not
make those credits available to NEW but only for FORWARDs ... there are
more details of course.
Hope this explains
Thanks
-Venky
On Fri, Dec 6, 2019 at 12:37 PM Mattias Rönnblom <
mattias.ronnb...@ericsson.com> wrote:
On 2019-12-06 17:32, Venky Venkatesh wrote:
Thanks Mattias for the clarifications.
1 more question: This time it is about the inflight accounting for DSW.
Here is my understanding: it seems to consider only the events which
are *inside
the scheduler* as in flight.
Yes, like all event devices, I believe.
I am trying to distinguish it from those which
have been currently given to cores by the scheduler. The latter are not
considered in flight since we dsw_port_return_credits as soon as
dsw_event_dequeue_burst.
A new dequeue means an implicit release of all unreleased events
dequeued in the previous call. It's standard Eventdev semantics.
So if these events which are in core currently do
a FORWARD, there is a chance that those can fail. Ideally those FORWARDs
should not fail -- which can happen with the current design as some NEWs
can hog those credits freed up by the ones which have been dequeued by
cores.
What you do to avoid this situation is set the new_event_threshold
low-enough, so NEW events don't block FORWARDed ones.
Is this design of DSW intentional or an omission? If it is an
omission I can work on a possible fix and run it by you.
This is not really a DSW design, but rather how Eventdev works.
