RE: [RFC PATCH] ring: adding TPAUSE instruction to ring dequeue

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

> -----Original Message-----
> From: Stephen Hemminger [mailto:stephen@networkplumber.org]
> 
> On Thu, 4 May 2023 16:11:31 +0000
> "Coyle, David" <david.coyle@intel.com> wrote:
> 
> > Hi Morten
> >
> > > -----Original Message-----
> > > From: Morten Brørup <mb@smartsharesystems.com>
> > >
> >
> > <snip>
> >
> > > Power saving is important for the environment (to save the planet and all
> > > that), so everyone should contribute, if they have a good solution. So
> even if
> > > our algorithm had a significant degree of innovation, we would probably
> > > choose to make it public anyway. Open sourcing it also makes it possible
> for
> > > chip vendors like Intel to fine tune it more than we can ourselves, which
> also
> > > comes back to benefit us. All products need some sort of power saving in
> to
> > > stay competitive, but power saving algorithms is not an area we want to
> > > pursue for competitive purposes in our products.
> > >
> > > Our algorithm is too simple to make a library at this point, but I have
> been
> > > thinking about how we can make it a generic library when it has matured
> > > some more. I will take your information about the many customers' need to
> > > have it invisibly injected into consideration in this regard.
> > >
> > > Our current algorithm works like this:
> > >
> > > while (running) {
> > > int more = 0;
> > > more += stage1();
> > > more += stage2();
> > > more += stage3();
> > > if (!more) sleep();
> > > }
> > >
> > > Each pipeline stage only returns 1 if it processed a full burst.
> Furthermore, if a
> > > pipeline stage processed a full burst, but happens to know that no more
> data
> > > is readily available for it, it returns 0 instead.
> > >
> > > Obviously, the sleep() duration must be short enough to avoid that the NIC
> > > RX descriptor rings overflow before the ingress pipeline stage is serviced
> > > again.
> > >
> > > Changing the algorithm to "more" (1 = more work expected by the pipeline
> > > stage) from "busy" (1 = some work done by the pipeline stage) has the
> > > consequence that sleep() is called more often, which has the follow-on
> > > consequence that the ingress stage is called less often, and thus more
> often
> > > has a full burst to process.
> > >
> > > We know from our in-house profiler that processing a full burst provides
> > > *much* higher execution efficiency (cycles/packet) than processing a few
> > > packets. This is public knowledge - after all, this is the whole point of
> DPDK's
> > > vector packet processing design! Nonetheless, it might surprise some
> people
> > > how much the efficiency (cycles/packet) increases when processing a full
> > > burst compared to processing just a few packets. I will leave it up to the
> > > readers to make their own experiments. :-)
> > >
> > > Our initial "busy" algorithm behaved like this:
> > > Process a few packets (at low efficiency), don't sleep, Process a few
> packets
> > > (at low efficiency), don't sleep, Process a few packets (at low
> efficiency),
> > > don't sleep, Process a few packets (at low efficiency), don't sleep,
> Process a
> > > few packets (at low efficiency), don't sleep, Process a few packets (at
> low
> > > efficiency), don't sleep, Process a few packets (at low efficiency), don't
> > > sleep, Process a few packets (at low efficiency), don't sleep, No packets
> to
> > > process (we are lucky this time!), sleep briefly, Repeat.
> > >
> > > So we switched to our "more" algorithm, which behaves like this:
> > > Process a few packets (at low efficiency), sleep briefly, Process a full
> burst of
> > > packets (at high efficiency), don't sleep, Repeat.
> > >
> > > Instead of processing e.g. 8 small bursts per sleep, we now process only 2
> > > bursts per sleep. And the big of the two bursts is processed at higher
> > > efficiency.
> > >
> > > We can improve this algorithm in some areas...
> > >
> > > E.g. some of our pipeline stages also know that they are not going to do
> > > anymore work for the next X amount of nanoseconds; but we don't use that
> > > information in our power management algorithm yet. The sleep duration
> > > could depend on this.
> > >
> > > Also, we don't use the CPU power management states yet. I assume that
> > > doing some work for 20 us at half clock speed is more power conserving
> than
> > > doing the same work at full speed for 10 us and then sleeping for 10 us.
> > > That's another potential improvement.
> > >
> > >
> > > What we need in generic a power management helper library are functions
> > > to feed it with the application's perception of how much work is being
> done,
> > > and functions to tell if we can sleep and/or if we should change the power
> > > management states of the individual CPU cores.
> > >
> > > Such a unified power management helper (or "busyness") library could
> > > perhaps also be fed with data directly from the drivers and libraries to
> > > support the customer use cases you described.
> >
> > [DC] Thank you for that detailed description, very interesting. There may
> > well be merit in upstreaming such an algorithm as a library once it has
> > matured as you said.
> >
> > Configuration could include specifying what a "full burst"
> > actually is. Different stages of a pipeline may also have different
> definitions
> > of busyness, so that may also need to considered:
> > - Some stages may perform an operation (e.g. running an acl rule check) on a
> > burst of packets and then it is complete

We don't have such a configuration. In our application framework, each pipeline stage reports back to the main loop if it expects to have more work readily available or not. So it is up to the pipeline stage to determine what it considers a "full burst" or "100 % busy".

Only the RX and TX stages actually compare the number of processed packets with their MAX_PKT_BURST equivalent. In addition to this, we also consider the return values from rte_eth_rx_queue_count() and/or the "available" output parameter from rte_ring_dequeue_burst() to determine if more data is readily available, or if the pipeline stage should return 0 ("no more work") even though it processed a full burst of packets.

> > - Other stages may be more asynchronous in nature e.g. enqueuing and
> > dequeuing to/from a crypto device or a QoS scheduler. The dequeue might
> > not dequeue any packets on a particular call of the dequeue API, but there
> > may still be packets waiting inside the crypto device or scheduler. Those
> waiting
> > packets would also need to be taken into account so as not to sleep for too
> long.

Good point!

This needs further consideration.

> >
> > Using such an API would require a workload developer to update their
> datapath
> > to report the pipeline stage busyness to the algorithm, but if those calls
> are
> > kept to a minimum, then that shouldn't be too much of a problem

Yes, that is what I am aiming at.

> >
> > Thanks,
> > David
> 
> I see two overlapping discussions here:
> 
> The first, is using some form of memory wait when or timed pause for the cases
> where it is spinning on contended region like lock or ring concurrency.
> There already is some of this available on Arm64 and having TPAUSE used on
> intel
> makes sense.  Using TPAUSE in rte_usleep is obvious good idea.

I haven't looked deeply into this. Hijacked the thread instead. :-)

> 
> The other is having some overall indication of busyness. This would be how
> often things like rx_burst and ring_dequeue get data to work on. A mechanism
> for this must be lightweight (ie per-core and minimum data collection),
> and plumbed into the telemetry system.

Agree.

> It makes sense that this would
> be a new DPDK EAL call that would be used in place of the sleep done
> by most applications in the main loop when not busy.

This would make it easy to plug in. Good point!

(Although EAL is bloated enough already, so I might prefer having it elsewhere.)

> Any solution should be architecture independent

Agree!

> 
> None of the designs presented so far seem complete and simple enough
> to be part of the main DPDK distribution. Keep working and experimenting.

It seems that we are quite a few organizations experimenting with this individually. And RFCs pop up on the list once in a while; but none of us have found the right solution yet.

That's why I appreciate open discussions like these, sharing experience and pointing out the pitfalls.

We just keep working, experimenting, and discussing. It will pay off one day.