Yafang Shao <laoar.shao@xxxxxxxxx> writes:
> On Fri, Jul 12, 2024 at 11:07 AM Huang, Ying <ying.huang@xxxxxxxxx> wrote:
>>
>> Yafang Shao <laoar.shao@xxxxxxxxx> writes:
>>
>> > On Fri, Jul 12, 2024 at 9:21 AM Huang, Ying <ying.huang@xxxxxxxxx> wrote:
>> >>
>> >> Yafang Shao <laoar.shao@xxxxxxxxx> writes:
>> >>
>> >> > On Thu, Jul 11, 2024 at 6:51 PM Huang, Ying <ying.huang@xxxxxxxxx> wrote:
>> >> >>
>> >> >> Yafang Shao <laoar.shao@xxxxxxxxx> writes:
>> >> >>
>> >> >> > On Thu, Jul 11, 2024 at 4:20 PM Huang, Ying <ying.huang@xxxxxxxxx> wrote:
>> >> >> >>
>> >> >> >> Yafang Shao <laoar.shao@xxxxxxxxx> writes:
>> >> >> >>
>> >> >> >> > On Thu, Jul 11, 2024 at 2:44 PM Huang, Ying <ying.huang@xxxxxxxxx> wrote:
>> >> >> >> >>
>> >> >> >> >> Yafang Shao <laoar.shao@xxxxxxxxx> writes:
>> >> >> >> >>
>> >> >> >> >> > On Wed, Jul 10, 2024 at 10:51 AM Huang, Ying <ying.huang@xxxxxxxxx> wrote:
>> >> >> >> >> >>
>> >> >> >> >> >> Yafang Shao <laoar.shao@xxxxxxxxx> writes:
>> >> >> >> >> >>
>> >> >> >> >> >> > The configuration parameter PCP_BATCH_SCALE_MAX poses challenges for
>> >> >> >> >> >> > quickly experimenting with specific workloads in a production environment,
>> >> >> >> >> >> > particularly when monitoring latency spikes caused by contention on the
>> >> >> >> >> >> > zone->lock. To address this, a new sysctl parameter vm.pcp_batch_scale_max
>> >> >> >> >> >> > is introduced as a more practical alternative.
>> >> >> >> >> >>
>> >> >> >> >> >> In general, I'm neutral to the change. I can understand that kernel
>> >> >> >> >> >> configuration isn't as flexible as sysctl knob. But, sysctl knob is ABI
>> >> >> >> >> >> too.
>> >> >> >> >> >>
>> >> >> >> >> >> > To ultimately mitigate the zone->lock contention issue, several suggestions
>> >> >> >> >> >> > have been proposed. One approach involves dividing large zones into multi
>> >> >> >> >> >> > smaller zones, as suggested by Matthew[0], while another entails splitting
>> >> >> >> >> >> > the zone->lock using a mechanism similar to memory arenas and shifting away
>> >> >> >> >> >> > from relying solely on zone_id to identify the range of free lists a
>> >> >> >> >> >> > particular page belongs to[1]. However, implementing these solutions is
>> >> >> >> >> >> > likely to necessitate a more extended development effort.
>> >> >> >> >> >>
>> >> >> >> >> >> Per my understanding, the change will hurt instead of improve zone->lock
>> >> >> >> >> >> contention. Instead, it will reduce page allocation/freeing latency.
>> >> >> >> >> >
>> >> >> >> >> > I'm quite perplexed by your recent comment. You introduced a
>> >> >> >> >> > configuration that has proven to be difficult to use, and you have
>> >> >> >> >> > been resistant to suggestions for modifying it to a more user-friendly
>> >> >> >> >> > and practical tuning approach. May I inquire about the rationale
>> >> >> >> >> > behind introducing this configuration in the beginning?
>> >> >> >> >>
>> >> >> >> >> Sorry, I don't understand your words. Do you need me to explain what is
>> >> >> >> >> "neutral"?
>> >> >> >> >
>> >> >> >> > No, thanks.
>> >> >> >> > After consulting with ChatGPT, I received a clear and comprehensive
>> >> >> >> > explanation of what "neutral" means, providing me with a better
>> >> >> >> > understanding of the concept.
>> >> >> >> >
>> >> >> >> > So, can you explain why you introduced it as a config in the beginning ?
>> >> >> >>
>> >> >> >> I think that I have explained it in the commit log of commit
>> >> >> >> 52166607ecc9 ("mm: restrict the pcp batch scale factor to avoid too long
>> >> >> >> latency"). Which introduces the config.
>> >> >> >
>> >> >> > What specifically are your expectations for how users should utilize
>> >> >> > this config in real production workload?
>> >> >> >
>> >> >> >>
>> >> >> >> Sysctl knob is ABI, which needs to be maintained forever. Can you
>> >> >> >> explain why you need it? Why cannot you use a fixed value after initial
>> >> >> >> experiments.
>> >> >> >
>> >> >> > Given the extensive scale of our production environment, with hundreds
>> >> >> > of thousands of servers, it begs the question: how do you propose we
>> >> >> > efficiently manage the various workloads that remain unaffected by the
>> >> >> > sysctl change implemented on just a few thousand servers? Is it
>> >> >> > feasible to expect us to recompile and release a new kernel for every
>> >> >> > instance where the default value falls short? Surely, there must be
>> >> >> > more practical and efficient approaches we can explore together to
>> >> >> > ensure optimal performance across all workloads.
>> >> >> >
>> >> >> > When making improvements or modifications, kindly ensure that they are
>> >> >> > not solely confined to a test or lab environment. It's vital to also
>> >> >> > consider the needs and requirements of our actual users, along with
>> >> >> > the diverse workloads they encounter in their daily operations.
>> >> >>
>> >> >> Have you found that your different systems requires different
>> >> >> CONFIG_PCP_BATCH_SCALE_MAX value already?
>> >> >
>> >> > For specific workloads that introduce latency, we set the value to 0.
>> >> > For other workloads, we keep it unchanged until we determine that the
>> >> > default value is also suboptimal. What is the issue with this
>> >> > approach?
>> >>
>> >> Firstly, this is a system wide configuration, not workload specific.
>> >> So, other workloads run on the same system will be impacted too. Will
>> >> you run one workload only on one system?
>> >
>> > It seems we're living on different planets. You're happily working in
>> > your lab environment, while I'm struggling with real-world production
>> > issues.
>> >
>> > For servers:
>> >
>> > Server 1 to 10,000: vm.pcp_batch_scale_max = 0
>> > Server 10,001 to 1,000,000: vm.pcp_batch_scale_max = 5
>> > Server 1,000,001 and beyond: Happy with all values
>> >
>> > Is this hard to understand?
>> >
>> > In other words:
>> >
>> > For applications:
>> >
>> > Application 1 to 10,000: vm.pcp_batch_scale_max = 0
>> > Application 10,001 to 1,000,000: vm.pcp_batch_scale_max = 5
>> > Application 1,000,001 and beyond: Happy with all values
>>
>> Good to know this. Thanks!
>>
>> >>
>> >> Secondly, we need some evidences to introduce a new system ABI. For
>> >> example, we need to use different configuration on different systems
>> >> otherwise some workloads will be hurt. Can you provide some evidences
>> >> to support your change? IMHO, it's not good enough to say I don't know
>> >> why I just don't want to change existing systems. If so, it may be
>> >> better to wait until you have more evidences.
>> >
>> > It seems the community encourages developers to experiment with their
>> > improvements in lab environments using meticulously designed test
>> > cases A, B, C, and as many others as they can imagine, ultimately
>> > obtaining perfect data. However, it discourages developers from
>> > directly addressing real-world workloads. Sigh.
>>
>> You cannot know whether your workloads benefit or hurt for the different
>> batch number and how in your production environment? If you cannot, how
>> do you decide which workload deploys on which system (with different
>> batch number configuration). If you can, can you provide such
>> information to support your patch?
>
> We leverage a meticulous selection of network metrics, particularly
> focusing on TcpExt indicators, to keep a close eye on application
> latency. This includes metrics such as TcpExt.TCPTimeouts,
> TcpExt.RetransSegs, TcpExt.DelayedACKLost, TcpExt.TCPSlowStartRetrans,
> TcpExt.TCPFastRetrans, TcpExt.TCPOFOQueue, and more.
>
> In instances where a problematic container terminates, we've noticed a
> sharp spike in TcpExt.TCPTimeouts, reaching over 40 occurrences per
> second, which serves as a clear indication that other applications are
> experiencing latency issues. By fine-tuning the vm.pcp_batch_scale_max
> parameter to 0, we've been able to drastically reduce the maximum
> frequency of these timeouts to less than one per second.
Thanks a lot for sharing this. I learned much from it!
> At present, we're selectively applying this adjustment to clusters
> that exclusively host the identified problematic applications, and
> we're closely monitoring their performance to ensure stability. To
> date, we've observed no network latency issues as a result of this
> change. However, we remain cautious about extending this optimization
> to other clusters, as the decision ultimately depends on a variety of
> factors.
>
> It's important to note that we're not eager to implement this change
> across our entire fleet, as we recognize the potential for unforeseen
> consequences. Instead, we're taking a cautious approach by initially
> applying it to a limited number of servers. This allows us to assess
> its impact and make informed decisions about whether or not to expand
> its use in the future.
So, you haven't observed any performance hurt yet. Right? If you
haven't, I suggest you to keep the patch in your downstream kernel for a
while. In the future, if you find the performance of some workloads
hurts because of the new batch number, you can repost the patch with the
supporting data. If in the end, the performance of more and more
workloads is good with the new batch number. You may consider to make 0
the default value :-)
> [0] 'Cluster' refers to a Kubernetes concept, where a single cluster
> comprises a specific group of servers designed to work in unison.
--
Best Regards,
Huang, Ying
