RE: [RFC] service: extend service function call statistics

["Morten Brørup" <mb@smartsharesystems.com>]

> From: Mattias Rönnblom [mailto:hofors@lysator.liu.se]
> Sent: Friday, 26 January 2024 09.28
> 
> On 2024-01-26 00:19, Morten Brørup wrote:
> >> From: Mattias Rönnblom [mailto:mattias.ronnblom@ericsson.com]
> >> Sent: Thursday, 25 January 2024 20.15
> >>
> >> Add two new per-service counters.
> >>
> >> RTE_SERVICE_ATTR_IDLE_CALL_COUNT tracks the number of service
> function
> >> invocations where no work was performed.
> >>
> >> RTE_SERVICE_ATTR_ERROR_CALL_COUNT tracks the number invocations
> >> resulting in an error.
> >>
> >> The semantics of RTE_SERVICE_ATTR_CALL_COUNT remains the same (i.e.,
> >> counting all invocations, regardless of return value).
> >>
> >> The new statistics may be useful for both debugging and profiling
> >> (e.g., calculate the average per-call processing latency for non-
> idle
> >> service calls).
> >>
> >> Service core tests are extended to cover the new counters, and
> >> coverage for RTE_SERVICE_ATTR_CALL_COUNT is improved.
> >
> > OK to all of the above. Good stuff.
> >
> >>
> >> The documentation for the CYCLES attributes are updated to reflect
> >> their actual semantics.
> >
> > If this is intended behavior, then updating the documentation seems
> appropriate - I would even go so far as considering it a bug fix.
> >
> > However, quite a few cycles may be consumed by a service before it
> can conclude that it had no work to do. Shouldn't that be considered
> time spent by the service? I.e. should the code be fixed instead of the
> documentation?
> >
> 
> Generally, polling for new work in the service is cheap, in my
> experience. But there's nothing in principle that prevents the
> situation
> your describe from occurring. You could add an "IDLE_CYCLES" counter in
> case that would ever be a real-world problem.
> 
> That wouldn't be a fix, but rather just returning to the old, subtly
> broken, (pre-22.11?) semantics.
> 
> Have a look at 809bd24 to see the rationale for the change. There's an
> example in 4689c57.
> 
> The cause of this ambiguity is due to the fact that the platform/OS
> (i.e., DPDK) doesn't know which service to "wake up" (which in turn is
> the result of the platform not being able to tracking input sources,
> like NIC RX queues or timer wheels) and thus must ask every service to
> check if it has something to do.

OK. Makes good sense.
So definitely fix the documentation, not the code. :-)

> 
> > Alternatively, keep the behavior (for backwards compatibility) and
> fix the documentation, as this patch does, and add an IDLE_CYCLES
> counter for time spent in idle calls.
> >
> > PS: We're not using DPDK service cores in our applications, so I'm
> not familiar with the details. We are using something somewhat similar
> (but homegrown), also for profiling and power management purposes, and
> my feedback is based on my experience with our own variant of service
> cores.
> >
> 
> When are you making the switch to service cores? :)

Our own "service cores" implementation has some slightly different properties, which we are still experimenting with.

E.g. in addition to the special return value "idle (no work to do)", we also have a special return value for "incomplete (more work urgently pending)" when a service processed a full burst and still has more work pending its input queue.

We are also considering returning a value to inform what time it needs to be called again. This concept is only an idea, and we haven't started experimenting with it yet.


From a high level perspective, the service cores library is much like an operating system's CPU scheduler, although based on voluntary time sharing. Many algorithms and many parameters can be considered. It can also tie into power management and prioritization of different tasks.

> 
> > Either way:
> >
> > Acked-by: Morten Brørup <mb@smartsharesystems.com>
> >
> 
> Thanks for the review.