On 2022-10-04 13:57, Bruce Richardson wrote:
> On Tue, Oct 04, 2022 at 11:15:19AM +0200, Morten Brørup wrote:
>>> From: Mattias Rönnblom [mailto:hof...@lysator.liu.se]
>>> Sent: Monday, 3 October 2022 22.02
>>
>> [...]
>>
>>> The functionality provided is very useful, and the implementation is
>>> clever in the way it doesn't require any application modifications.
>>> But,
>>> a clever, useful brittle hack is still a brittle hack.
>>>
>
> I think that may be a little harsh here. After all, this is a feature which
> is build-time disabled and runtime disabled by default, so like many other
> components it's designed for use when it makes sense to do so.
>
So you don't think it's a hack? The driver level and the level of basic
data structures (e.g., the ring) is the appropriate level to classify
cycles into useful and not useful? And you don't think all the shaky
assumptions makes it brittle?
Runtime configurable or not doesn't make a difference in this regard, in
my opinion. On the source code level, this code is there, and making it
compile-time conditional just makes matters worse.
Had this feature been limited to a small library, it would made a
difference, but it's smeared across a wide range of APIs, and this list
is not yet complete. Anything than can produce items of work need to be
adapted.
That said, it's not obvious how this should be done. The higher-layer
constructs where this should really be done aren't there in DPDK, at
least not yet.
Have you considered the option to instrument rte_pause()? It's the
closes DPDK has to the (now largely extinct) idle loop in an OS kernel.
It too would be a hack, but maybe a less intrusive one.
> Furthermore, I'd just like to point out that the authors, when doing the
> patches, have left in the hooks so that even apps, for which the "for-free"
> scheme doesn't work, can still leverage the infrastructure to have the app
> itself report the busy/free metrics.
>
If this is done properly, in a way that the data can reasonably be
trusted and it can be enabled in runtime without much of a performance
implication, tracking lcore load could be much more useful, than just
best effort-telemetry.
Why is it so important not to require changes to the application? The
changes are likely trivial, not unlike those I've submitted for the
equivalent bookkeeping for DPDK services.
>>> What if there was instead a busyness module, where the application
>>> would
>>> explicitly report what it was up to. The new library would hook up to
>>> telemetry just like this patchset does, plus provide an explicit API to
>>> retrieve lcore thread load.
>>>
>>> The service cores framework (fancy name for rte_service.c) could also
>>> call the lcore load tracking module, provided all services properly
>>> reported back on whether or not they were doing anything useful with
>>> the
>>> cycles they just spent.
>>>
>>> The metrics of such a load tracking module could potentially be used by
>>> other modules in DPDK, or by the application. It could potentially be
>>> used for dynamic load balancing of service core services, or for power
>>> management (e.g, DVFS), or for a potential future deferred-work type
>>> mechanism more sophisticated than current rte_service, or some green
>>> threads/coroutines/fiber thingy. The DSW event device could also use it
>>> to replace its current internal load estimation scheme.
>>
>> [...]
>>
>> I agree 100 % with everything Mattias wrote above, and I would like to voice
>> my opinion too.
>>
>> This patch is full of preconditions and assumptions. Its only true advantage
>> (vs. a generic load tracking library) is that it doesn't require any
>> application modifications, and thus can be deployed with zero effort.
>>
>> I my opinion, it would be much better with a well designed generic load
>> tracking library, to be called from the application, so it gets correct
>> information about what the lcores spend their cycles doing. And as Mattias
>> mentions: With the appropriate API for consumption of the collected data, it
>> could also provide actionable statistics for use by the application itself,
>> not just telemetry. ("Actionable statistics": Statistics that is directly
>> usable for decision making.)
>>
>> There is also the aspect of time-to-benefit: This patch immediately provides
>> benefits (to the users of the DPDK applications that meet the
>> preconditions/assumptions of the patch), while a generic load tracking
>> library will take years to get integrated into applications before it
>> provides benefits (to the users of the DPDK applications that use the new
>> library).
>>
>> So, we should ask ourselves: Do we want an application-specific solution
>> with a short time-to-benefit, or a generic solution with a long
>> time-to-benefit? (I use the term "application specific" because not all
>> applications can be tweaked to provide meaningful data with this patch. You
>> might also label a generic library "application specific", because it
>> requires that the application uses the library - however that is a common
>> requirement of all DPDK libraries.)
>>
>> Furthermore, if the proposed patch is primarily for the benefit of OVS, I
>> suppose that calls to a generic load tracking library could be added to OVS
>> within a relatively short time frame (although not as quick as this patch).
>>
>> I guess that the developers of this patch initially thought that it was
>> generic and usable for the majority of applications, and it came as somewhat
>> a surprise that it wasn't as generic as expected. The DPDK community has a
>> good review process with open discussions and sharing of thoughts and ideas.
>> Sometimes, an idea doesn't fly, because the corner cases turn out to be more
>> common than expected. I'm sorry to say it, but I think that is the case for
>> this patch. :-(
>>
>
> I'd actually like to question this last statement a little.
>
> I think we in the DPDK community are very good at coming up with
> theoretical examples where things don't work, but are they really cases
> that occur commonly in the real-world?
>
> I accept, for example, that the "for free" approach would not be suitable
> for something like VPP which does multiple polls to gather packets before
> processing, but for some of the other cases I'd question their commonality.
> For example, a number of objections have focused on the case where
> allocation of buffers fails and so the busyness gets counted wrongly. Are
> there really (many) apps out there where running out of buffers is not a
> much more serious problem than incorrectly reported busyness stats?
>
Many, if not all, non-trivial DPDK applications will poll multiple
sources of work, some of which almost always will fail to produce any
items. In such cases, they will transit between the busy and idle state,
potentially several times, for every iteration in their lcore thread
poll loop. That will cause a performance degradation if this features is
used, and there's nothing they can do to fix it from the application
level, assuming they find this telemetry statistic useful and don't want
it disabled. So, not "for free", although may be you can still argue
it's a bargain. :)
> I'd also say that, in my experience, the non-open-source end-user apps tend
> very much to use DPDK based on the style of operation given in our DPDK
> examples, rather than trying out new or different ways of working. (Maybe
> others have different experiences, though, and can comment). I also tend to
> believe that open-source software using DPDK probably shows more variety in
> how things are done, which is not representative of a lot of non-OSS users
> of DPDK.
>
> Regards,
> /Bruce
