Re: [E-impact] kWh/GB should be banned and the IETF should warn against its use

[Hesham ElBakoury <[e-mail]>]

In the 2022 HotCarbon conference a paper from Max Planck Institute
presented a paper on "*Odd One Out: Energy is not like Other Metrics*" [
https://hotcarbon.org/pdf/hotcarbon22-anand.pdf]

The authors posit that "energy cannot be treated as just another metric".
They identify 4 factors that distinguish
energy from other typical metrics used today: "(1) energy consumption is
all-encompassing, i.e., every aspect of a system from software to hardware
consumes energy; (2) the same
computation running on different hardware, platforms, or environments, can
vary substantially in its energy usage; (3) the motivation behind measuring
energy consumption is different, as it is neither performance nor
correctness driven; and (4
energy cannot be measured at extremely fine granularities like, e.g., the
performance counters of modern system"

Although the paper focused on data center network but the same reasoning
applies to other networks.

Paper abstract says: "Energy requirements for datacenters are growing at a
fast
pace. Existing techniques for making datacenters efficient focus on
hardware. However, the gain in energy efficiency
that can be achieved without making the applications energy-aware is
limited. To overcome this limitation, recent work has
proposed making the software running in datacenters energy aware. To do so,
we must be able to track energy consumption
at various granularities at the software level – (i) process level; (ii)
application level; (iii) end-to-end request level. Currently, existing
software energy-tracking techniques primarily focus on tracking energy at
the process or application
level; only a few techniques track energy at an end-to-end request level.
However, not tracking energy at an end-to-end request level can lead to
false software optimizations and cause a decrease in energy efficiency."

The authors conclude that "While hardware meters are not able to measure
end-to-end request energy consumption, their deployment can help the tracer
to get more accurate data. Furthermore, to encourage application developers
to treat
energy as a first-class citizen, cloud providers can shift towards
energy-consumption based pricing. Finally, we can involve service
customers: as more peo-
ple are committed to reduce carbon emissions, the energy consumption of a
request may be of increasing relevance to
customers. We envision services displaying to end users the
energy-consumption estimates of their requests, along the lines of flight
booking services, which annotate each flight option with their estimated
carbon emissions to encourage
customers to make a more environment-friendly choice."

Hesham

On Mon, May 15, 2023, 12:29 PM Toerless Eckert <[e-mail]> wrote:

> I think we really should avoid to be dragged into the low layer metric
> alone, as
> much fun as they are, given how we easily understand them.
>
> IMHO the corect question is to ask how we want to finance the Internet
> infrastructure
> while simultaneously driving down its energy consumption. That is the
> ultimate problem
> to solve. All other work items such as metrics have to trickle down from
> that.
>
> For example, if i have three internet edge providers putting up competing
> infrastructures,
> none of them ever getting to max utilization, we know we're wasting
> energy. How do we capture that
> as a metric ? And how do we weigh that against competitiveness and
> innovation ?
>
> If we have three competing core providers, today they likely all have high
> utilization,
> so with some good will, this is better competition, because they all add
> necessary
> transit bandwidth. How do we capture that ?
>
> Then we have the problem of peak vs. average consumption. As long as an
> operator can
> afford it, capacity will be built out for peak - and cost and energy
> consumption will
> be for that peak. And if minimum energy consumption of equipment is maybe
> just 20%
> lower than for peak performance, but average utilization of the equipment
> is 80% lower
> than peak, then we have a big mismatch/problem of energy waste. How do we
> capture this
> in metrics ?
>
> Sure, we can start doing instrumentation of low-level power consumption
> metrics in equipment,
> and we should. But by itself, this work will not move the needle when
> nobody looks at
> the broader picture and those higher layer metrics. It is only those
> higher layer metrics
> that will cause actionable progress IMHO. Peak to average utilization for
> example. And
> key contributors to peak utilization as the worst offenders of energy
> waste by need ot
> building out against such peaks.
>
> Cheers
>     Toerless
>
> Disclaimer: Having worked for 30+ years on multicast, i do of course have
> some opinion
> on peak shaving, so when i see how currend trends in streaming are often
> ignoring many of
> those possibilities and i am saying that it would be nice to have metrics
> to better expose
> the network energy waste of such applications - then of course i am not
> talking out of
> a vacuum or purely from a "save the planet" perspective, but also from the
> perspective of
> creating more awareness for such helpful technologies.
>
>
> On Mon, May 15, 2023 at 03:25:40PM +0000, Rob Wilton (rwilton) wrote:
> > Hi Rudolf,
> >
> > I struggle to understand how “traffic and energy use are in no way and
> have never been related”.
> >
> > On some modern forwarding ASICs, between a third and quarter of the
> power for the forwarding ASIC comes from the dynamic load (e.g., packets
> switched per-second and bytes switched per-second).  If you consider the
> other power usage, e.g., optics, fabric, and fan, then the overall dynamic
> power difference for the device will be much smaller.  But I think that it
> is probably a somewhat inaccurate generalization to say that they are not
> related at all, at least for some modern devices.  Clearly, much more
> significant power savings would be available by disabling optics or
> shutting down linecards, which may allow some of the fabric planes to be
> taken offline, fan speeds to be reduced, etc.
> >
> > Another way of potentially looking at this is to figure out the total
> power usage of a network (e.g., perhaps your local ISP), and figure out
> what portion of the traffic the user is consuming at peak times (i.e.,
> since that it probably the driving factor for network capability
> planning).  This might give some indirect indication as to what proportion
> of the network energy usage is effectively due to a given user.  I
> appreciate that the overall figure may be very crude, but providing an
> end-user with some rough indication is possibly better than providing none
> at all.
> >
> > Regards,
> > Rob
> >
> >
> > From: E-impact <[e-mail]> On Behalf Of Rudolf van der
> Berg
> > Sent: 12 May 2023 20:05
> > To: Hesham ElBakoury <[e-mail]>
> > Cc: Romain Jacob <[e-mail]>; [e-mail]
> > Subject: Re: [E-impact] kWh/GB should be banned and the IETF should warn
> against its use
> >
> > Hesham,
> >
> > Look at the title of the thread. The kWh/GB metic is nonsense. It
> doesn't convey any comparable and actionable meaning. The numbers are
> incomparable because traffic and energy use are in no way and have never
> been related. How is it possible that with the exact same network and
> equipment the energy use is stable when traffic goes up? Simple it is a
> useless unscientific metric.
> >
> > Op vr 12 mei 2023 20:48 schreef Hesham ElBakoury <[e-mail]
> <mailto:[e-mail]>>:
> > The DT CR 2022 reports 91 kwh/terabyte. This is higher than 60
> kwh/terabyte. Is this because the 60kwh is for the mobile network and the
> remaining 31kwh is for the DT wired network?
> >
> > If DT uses .06 kwh/GB for mobile network,  that is better than the
> 0.1-0.6 kwh/GB range reported by GSMA. Does that mean that DT network
> energy efficiency is high or ... ?
> >
> > Thanks,
> > Hesham
> >
> >
> > On Fri, May 12, 2023, 8:22 AM Rudolf van der Berg <
> [e-mail]<mailto:[e-mail]>> wrote:
> > GSMA is mobile only..
> >
> > Op vr 12 mei 2023 17:18 schreef Hesham ElBakoury <[e-mail]
> <mailto:[e-mail]>>:
> > Thanks Rudolf!
> > For DT the numbers are getting smaller every year. I would expect that
> the numbers in 2023 to be smaller than 2022.
> >
> > In a different email you mentioned that Deutsche Telekom reports that it
> uses 60kWh per Terabyte.  This is different from the numbers reported in
> 2022 CR report.
> >
> > You mentioned that of the participating operators, in this year GSMA
> report  "66% performed in the 0.1–0.6 kWh/GB range  meaning that most of
> the included networks used 0.1–0.6 kWh to transfer 1 GB of data. Values
> outside of this range can be explained by inefficient networks, consumer
> habits or low data traffic per connection". DT 2022 CR report numbers are
> smaller than 0.1 kwh/GB. How to explain this low value?
> >
> > Hesham
> >
> > On Fri, May 12, 2023, 7:32 AM Rudolf van der Berg <
> [e-mail]<mailto:[e-mail]>> wrote:
> > Here is annual report 2022
> https://report.telekom.com/annual-report-2022/services/downloads.html
> >
> https://www.cr-report.telekom.com/2022/management-facts/environment/energy-consumption-efficiency#atn-19726-19728,atn-19726-19729,atn-19726-19730
> >
> >
> > Op vr 12 mei 2023 16:19 schreef Hesham ElBakoury <[e-mail]
> <mailto:[e-mail]>>:
> > The DT CR-2021 report describes the value for kwh/terabyte [
> https://www.cr-report.telekom.com/2021/management-facts/environment/energy-consumption-efficiency#atn-19726-19729,atn-19726-19730,atn-19726-19733,atn-19726-19728
> ]
> >
> > Note that the total energy consumption increases by 4% year over year
> while the kwh/terabit decreases year over year.
> >
> > I am not sure how the numbers in kwh/GB which are reported in this email
> thread are calculated?
> >
> > Hesham
> >
> >
> >
> > On Fri, May 12, 2023, 2:20 AM Romain Jacob <[e-mail]<mailto:
> [e-mail]>> wrote:
> >
> >
> > So no DT is not overprovisioned and utilisation is actively obstructed.
> Or do you have other data and sources that support your assessment?
> >
> > Not at all. It was not an assessment, only a suspicion. I only mentioned
> DT because I understood your previous email as suggesting that DT was maybe
> not to blame for their poor kWh/GB numbers -- it seems I understood wrong
> and DT is indeed to blame. I have no issues with that.
> >
> > --
> > Romain JACOB
> > Postdoctoral Researcher
> > ETH Zurich
> > Networked Systems Group (NSG)
> > Lead: Prof. Laurent Vanbever
> > www.romainjacob.net<https://www.romainjacob.net/>
> > @RJacobPartner<https://twitter.com/RJacobPartner>
> > @[e-mail]<https://discuss.systems/@jacobr>
> > Gloriastrasse 35, ETZ G81
> > 8092 Zurich
> > [phone]
> > --
> > E-impact mailing list
> > [e-mail]<mailto:[e-mail]>
> > https://www.ietf.org/mailman/listinfo/e-impact
>
> > --
> > E-impact mailing list
> > [e-mail]
> > https://www.ietf.org/mailman/listinfo/e-impact
>
>
> --
> ---
> [e-mail]
>