3 ms·

I have only read the abstract, so maybe I'm missing something, but how would moving workloads to different times really lower CO2 in the long term unless it's u

I have only read the abstract, so maybe I'm missing something, but how would moving workloads to different times really lower CO2 in the long term unless it's using surplus green power? They talk about "less carbon-intensive" supply, but if it's still 1% coal power at a particular time then using more power at that time will require burning coal to satisfy demands? It would seem to only help to place workloads at times where it could use up otherwise unused solar or wind power. Also, leaving hardware idle is both having a CO2 cost, too (production of hardware that's not used), and, other people who are not CO2 conscious will probably move their workloads to those times where the hardware is under utilized and hence cheaper. So, the whole thing is a zero sum game, making one company's workloads appear greener at the cost of making other companies' workloads worse? (Edit: removed negative tone.)

birnbaum 5y ago
Co-author here :) Regarding your first point: You are talking about so-called marginal carbon intensity - and you are right, this would indeed be a better metric for incentivizing scheduling. Unfortunately it is extremely hard to compute (or better: to estimate), we discuss this in Section 3.4 of the paper. Regarding the second point: This paper is more written from a data center operator or service provider perspective. They could, for example, (financially?) incentivize their users to give them some slack in when to schedule workloads. Simply put: Computing should be expensive when energy is dirty. Governments could incentivize this via carbon pricing mechanisms (that are actually currently discussed at COP26)
- pflanze 5y ago
  In section 3.4 you end with "we consider marginal carbon intensity to be no practical signal for demand management at this point due to high uncertainties". But for the direct actual ecological value that would be the only number that counts, according to my understanding, so basically you're agreeing that right now temporal shifts have no ecological benefit? It appears that you hope that in the future, when marginal carbon intensity data might become available, and temporal shifting actually has value, this is a useful approach, and that as of today, preparing for that future is worthwhile. I can see that point. I think this should be pointed out--this changes the proposal from something that looks like pseudo actionism (that I'm sure some companies' PR departments will still love) to an actual valid plan, even if it doesn't have immediate benefits. PS. also, the marginal carbon intensity benefit still has to outweight the CO2 cost for buying more hardware (to afford running the jobs at those time points, and consequentially leaving the hardware unused at the corresponding counter times). Have you looked into this?
  - birnbaum 5y ago
    That was not was I was trying to say: Overall, we _should_ try to shift workload towards times where the average carbon intensity is low. Statements like "energy is usually cleaner at 1pm than at 7pm because of solar power" absolutely hold true. It's just that from the perspective of shifting a single compute job, marginal carbon intensity could be a better metric. However, imho that's super speculative: The question which power plant will produce the "next" kWh of energy is very complex. It depends on a variety of factors like energy prices, weather (forecasts), expected future demand, contracts between energy providers and grid operators, etc. Additionally, one also needs information for all neighboring regions, as fluctuations in supply and demand are often regulated via cross-regional flows. That means, also in the future we will most likely not have a perfect methodology for estimating marginal carbon intensity. Regarding the hardware: No, we did not look into this in this paper. The goal should not be to buy more hardware and data centers are barely ever used at full capacity. Also flexible workloads are not only important for carbon footprint (think of Spot VMs/Instances). However, real life resource and business constrains will definitely lessen the potential of savings.
    - pflanze 5y ago
      > That was not was I was trying to say Maybe, but maybe it should be what you should say? I'm consciously trying to nudge you into this direction. > Statements like "energy is usually cleaner at 1pm than at 7pm because of solar power" absolutely hold true. Sure. But is using more energy at those times also cleaner? I provided for a logic argument that it is only if there's surplus green energy at that time, i.e. if the need doesn't have to be satisfied by increasing use of non green energy sources during those times. After which you led me to the term of "marginal carbon intensity" which seems to express exactly this, which I'm happy with, so I said then that's the only number that matters, and if you don't have that number then currently you can't draw conclusion about the whole being greener through the shifting. There's an argument to be made that one should move loads to 1pm since then the country could increase build out of solar energy, since there will be users of the surplus (instead of it needing storage to absorb, or left unused, or being used by more dubious consumers)! That would again be a good argument for the future, and I agree with it. But that's an argument I haven't seen you make--maybe you should. > The goal should not be to buy more hardware and data centers are barely ever used at full capacity Iff data centers are normally lesser used during times where the marginal carbon intensity is low, then yes, cool, no new hardware needs to be bought. Outside of that, of course you could just ignore the need for more hardware and argue that you only shift so little load that it is insignificant compared to the reserve. But that would not be a plan for serious changes, just free riding? Maybe it's the same thinking that leads you (and others, you cite Google in the paper) used to argue that "at 1pm the mix is cleaner, hence use energy at that time": if the amount of load shift is very small, you are tempted to argue that the mix doesn't change, hence the shift is beneficial. But again, this is just ignoring an effect, since evidently if you shift a lot then that won't be true any more. I'm sure it's tempting to argue away the effect, but that's what greenwashing tries to do in general, and may be what makes a lot of people in this discussion angry. So better distinguish it from greenwashing as well as you can. P.S.: just using up the reserve might of course be fine if moving away the shifted jobs again in case there's high non-shiftable demand coming up works quickly enough. (But it will be interesting to know how much that would be.) I don't really want to tear down your research, and some kind of simulation package to find out how much potential there is (and under what assumptions for the future etc.) would surely be cool. And I guess I've reached the point where I have to read the rest of the paper :)

birnbaum 5y ago

Co-author here :) Regarding your first point: You are talking about so-called marginal carbon intensity - and you are right, this would indeed be a better metric for incentivizing scheduling. Unfortunately it is extremely hard to compute (or better: to estimate), we discuss this in Section 3.4 of the paper. Regarding the second point: This paper is more written from a data center operator or service provider perspective. They could, for example, (financially?) incentivize their users to give them some slack in when to schedule workloads. Simply put: Computing should be expensive when energy is dirty. Governments could incentivize this via carbon pricing mechanisms (that are actually currently discussed at COP26)

pflanze 5y ago
In section 3.4 you end with "we consider marginal carbon intensity to be no practical signal for demand management at this point due to high uncertainties". But for the direct actual ecological value that would be the only number that counts, according to my understanding, so basically you're agreeing that right now temporal shifts have no ecological benefit? It appears that you hope that in the future, when marginal carbon intensity data might become available, and temporal shifting actually has value, this is a useful approach, and that as of today, preparing for that future is worthwhile. I can see that point. I think this should be pointed out--this changes the proposal from something that looks like pseudo actionism (that I'm sure some companies' PR departments will still love) to an actual valid plan, even if it doesn't have immediate benefits. PS. also, the marginal carbon intensity benefit still has to outweight the CO2 cost for buying more hardware (to afford running the jobs at those time points, and consequentially leaving the hardware unused at the corresponding counter times). Have you looked into this?
- birnbaum 5y ago
  That was not was I was trying to say: Overall, we _should_ try to shift workload towards times where the average carbon intensity is low. Statements like "energy is usually cleaner at 1pm than at 7pm because of solar power" absolutely hold true. It's just that from the perspective of shifting a single compute job, marginal carbon intensity could be a better metric. However, imho that's super speculative: The question which power plant will produce the "next" kWh of energy is very complex. It depends on a variety of factors like energy prices, weather (forecasts), expected future demand, contracts between energy providers and grid operators, etc. Additionally, one also needs information for all neighboring regions, as fluctuations in supply and demand are often regulated via cross-regional flows. That means, also in the future we will most likely not have a perfect methodology for estimating marginal carbon intensity. Regarding the hardware: No, we did not look into this in this paper. The goal should not be to buy more hardware and data centers are barely ever used at full capacity. Also flexible workloads are not only important for carbon footprint (think of Spot VMs/Instances). However, real life resource and business constrains will definitely lessen the potential of savings.
  - pflanze 5y ago
    > That was not was I was trying to say Maybe, but maybe it should be what you should say? I'm consciously trying to nudge you into this direction. > Statements like "energy is usually cleaner at 1pm than at 7pm because of solar power" absolutely hold true. Sure. But is using more energy at those times also cleaner? I provided for a logic argument that it is only if there's surplus green energy at that time, i.e. if the need doesn't have to be satisfied by increasing use of non green energy sources during those times. After which you led me to the term of "marginal carbon intensity" which seems to express exactly this, which I'm happy with, so I said then that's the only number that matters, and if you don't have that number then currently you can't draw conclusion about the whole being greener through the shifting. There's an argument to be made that one should move loads to 1pm since then the country could increase build out of solar energy, since there will be users of the surplus (instead of it needing storage to absorb, or left unused, or being used by more dubious consumers)! That would again be a good argument for the future, and I agree with it. But that's an argument I haven't seen you make--maybe you should. > The goal should not be to buy more hardware and data centers are barely ever used at full capacity Iff data centers are normally lesser used during times where the marginal carbon intensity is low, then yes, cool, no new hardware needs to be bought. Outside of that, of course you could just ignore the need for more hardware and argue that you only shift so little load that it is insignificant compared to the reserve. But that would not be a plan for serious changes, just free riding? Maybe it's the same thinking that leads you (and others, you cite Google in the paper) used to argue that "at 1pm the mix is cleaner, hence use energy at that time": if the amount of load shift is very small, you are tempted to argue that the mix doesn't change, hence the shift is beneficial. But again, this is just ignoring an effect, since evidently if you shift a lot then that won't be true any more. I'm sure it's tempting to argue away the effect, but that's what greenwashing tries to do in general, and may be what makes a lot of people in this discussion angry. So better distinguish it from greenwashing as well as you can. P.S.: just using up the reserve might of course be fine if moving away the shifted jobs again in case there's high non-shiftable demand coming up works quickly enough. (But it will be interesting to know how much that would be.) I don't really want to tear down your research, and some kind of simulation package to find out how much potential there is (and under what assumptions for the future etc.) would surely be cool. And I guess I've reached the point where I have to read the rest of the paper :)