Beyond \COtEmissions: The Overlooked Impact of Water Consumption of Information Retrieval Models

Concerns about energy consumption in the broader field of artificial intelligence (AI) were heightened after the study from Strubell et al. (2019), which was one of the first to highlight the high energy usage and emissions produced by large language models. This study and others before it (Albers, 2010; Belkhir and Elmeligi, 2018) have fueled a growing interest in measuring the energy consumption of research in related fields such as natural language processing (NLP) and machine learning (ML). Within the field of IR, the environmental impacts of technology at scale have been a concern for at least a decade (Chowdhury, 2012).

In general, there are two approaches one can take to quantifying energy and emissions: Life Cycle Assessment (LCA) (Finnveden et al., 2009) and power consumption measurement. The ISO standard defines LCA as the “compilation and evaluation of the inputs, outputs and the potential environmental impacts of a product system throughout its life cycle” (5, 2006). Due to its complexity and the number of resources required (Chowdhury, 2012), most studies choose to measure power consumption directly. Indeed, since then, there have been many efforts to measure energy and emissions for IR systems (Catena, 2015; Catena and Tonellotto, 2015; Catena et al., 2015; Catena and Tonellotto, 2017; Catena et al., 2018; Blanco et al., 2016). Given the explosion in neural models for search based on transformers (Dai and Callan, 2019; Hofstätter et al., 2020; MacAvaney et al., 2020; Zhuang and Zuccon, 2021b; Qu et al., 2021; Zhuang and Zuccon, 2021a; Xiong et al., 2020; Formal et al., 2021; Mallia et al., 2021; Lin and Ma, 2021), quantifying the energy and emissions of these systems is more critical than ever.

Recently, Scells et al. (2022) proposed a framework for minimizing energy usage and emissions for IR with their ‘reduce, reuse, recycle’ methodology. For each concept, they outlined several ways IR practitioners can lower their energy usage: for example, one can reduce the number of experiments they do; reuse existing pre-trained models for experiments; or recycle existing pre-trained models that were initially trained on one task but apply them to another task. Like many other studies before it, this research focused on energy consumption. In this paper, we extend this methodology to consider not only the energy consumption (and, by extension, emissions) but also the water consumption. To the best of our knowledge, our paper marks the first study to investigate the water consumption of IR systems, and neural IR models in particular.

Water consumption in data centers predominantly occurs due to two distinct factors: first, as an indirect consequence of generating electricity, typically through thermoelectric power sources, and second, as a direct requirement for cooling systems that help maintain the ideal operating environment. These factors are pictured in Figure 1. While water use for electricity generation is well known, its usage for data center cooling is likely less known to IR researchers.

The increase in model complexity associated with the latest advancement of AI models has seen the need for increasingly powerful servers. High-performance servers generate significant amounts of heat, which must be dissipated to maintain optimal operating conditions. Traditional cooling systems often rely on water, using evaporative cooling towers or chilled water systems to maintain temperatures. Consequently, the water consumption of these data centers can be significant (Mytton, 2021). For example, Google has reported that its data centers consumed 11.4 billion liters of water in the financial year 2017 and 15.8 billion in FY2018¹¹1https://services.google.com/fh/files/misc/google_2019-environmental-report.pdf.. The water used to cool data centers is often potable, thus reducing the amount of drinking water available to the population.

The type of cooling system plays a crucial role in determining water usage. For instance, air-cooled systems typically consume less water than water-cooled systems but may be less efficient at dissipating heat. Moreover, the local climate can also impact water consumption; data centers in hot and arid regions may require more water for cooling than those in cooler climates (Karimi et al., 2022). Finally, seasons and time of day also impact water consumption related to data centers cooling. The water used in data centers’ cooling towers is consumed in two ways:

• •

  through evaporation, which occurs as part of the process of cooling, where hot water returning from the data center is sprayed through water distribution nozzles across a cooling fill and then collected at the bottom of the cooling tower in a cold water basis, from where water is pumped back into the chiller connected to the data center’s air conditioning; and

• •

  through the process of blow down, where the water in the pipelines of the data center is flushed. This process of draining water is required to reduce salt, impurity, algae and bacteria concentration, which can cause damage to the cooling system. The higher the water quality, the less blow down of water.

Related to the blow down process is the concept of cycles of concentration $S$: the number of times the dissolved minerals and salts in the circulating water are concentrated compared to the concentration in the makeup water. This concept represents how efficiently the cooling tower system uses and recycles water by measuring how much water is evaporated and concentrated before discharge. A higher $S$ value indicates that the cooling tower is more efficient in reusing water, reducing overall water consumption and discharge. However, note that as the cycles of concentration increase, so does the concentration of dissolved solids, which can lead to scaling, corrosion, and other issues within the cooling system.

Water consumption related to data centers operations has attracted increasing attention from the research community (Mytton, 2021; Brocklehurst, 2021; Karimi et al., 2022), also in the context of complex and computationally-demanding AI models (Weidinger et al., 2022; Li et al., 2023). In particular, Li et al. (2023) provide a framework to estimate the water consumption of AI models and investigate the water consumption associated with the training of large language models such as LaMDA (Thoppilan et al., 2022). We build on the framework of Li et al. (2023) to investigate the water consumption related to several common IR models.

We build upon the framework for measuring the water efficiency of AI models set by Li et al. (2023). In this framework, the water consumption $W$ of a model $\mathcal{M}$ is measured as the sum of the water consumed for the cooling of the data center (on-site water consumption, $W_{on}(\mathcal{M})$) and of the water consumed for the production of the electricity used by the data center (off-site water consumption, $W_{off}(\mathcal{M})$):

The on-site water consumption can then be calculated with respect to the energy $e(\mathcal{M})$ used by the IR model and the water usage effectiveness $WUE_{on}$ of the data center. These can be parameterized by the time energy and water usage occur. This is because water consumption has daytime and season dependencies. Recall that the water is used to cool the data center. In times of the day that are hotter, e.g. in the early afternoons as opposed to the early mornings, or the hotter seasons, e.g., summer as opposed to winter, water consumption will be higher.

We account for this when computing $WUE_{on}$, which then is dependent from $e(t,\mathcal{M})$ and $WUE_{on}(t)$, where $t$ represents a specific time interval (e.g., these quantities could be computed for a 15 minutes interval). Note furthermore that the power consumption of a model may not be constant across all time intervals (e.g., $e(t=i,\mathcal{M})\neq e(t=j,\mathcal{M})$): this may be the case for example for a model that uses primarily CPU computing in a time interval and GPU computing in another time interval. Given that the water usage effectiveness $WUE_{on}$ of the data center depends on time, running different parts of an IR model pipeline, e.g., different times of the day, may give rise to different water utilization profiles. Given this, the on-site water consumption is calculated as follows:

Water usage effectiveness $WUE_{on}(t)$ typically depends on the cycles of concentration $S$ associated with the blow down of water used in the cooling tower and the outside wet-bulb temperature $T_{w}$.

The cycles of concentration for a cooling tower depend on the actual cooling tower specifications and the water quality. As detailed in Section 2.2, high-quality water, i.e. with few residues and impurities, requires fewer blow downs of the cooling tower’s pipelines. For example, we were able to acquire the cycles of concentration required by two cooling towers of different brands and installed in different locations of the same city in Brisbane, Australia – cooling tower $A$, located at a private organization, had $S_{A}=2.25$, while cooling tower $B$, located at a public hospital, had $S_{B}=1.33$.

The wet-bulb temperature is the temperature that is measured by a thermometer exposed to the air, and its bulb is covered with a wet wick, which is then exposed to moving air. As water evaporates from the wick, it cools the thermometer bulb, and the temperature reading reflects the cooling effect. Note that also $T_{w}$ is dependent on time (of day and season) as it is influenced by factors such as the air temperature, humidity, air pressure, and wind speed. Thus it can be parameterized accordingly, i.e. $T_{w}(t)$. For example, for Brisbane where the previous two cooling towers are located, the mean annual 9am $T_{w}(t=9am)=63.5F$ (min: 53.6F, max: 71.4F), while the mean 3pm $T_{w}(t=3pm)=65.3F$ (min: 57F, max: 72.3F).

We adopt the same modeling of a cooling tower used by Li et al. (2023) to compute $WUE_{on}(t)$:

Next we consider how to compute the off-site water consumption $W_{off}(\mathcal{M})$. The off-site water consumption is related to the cooling of the power plant, e.g., in the case of a nuclear or coal power station, and/or the actual production of the electricity, e.g., in the case of a hydroelectric power station. The off-site water consumption could also be null for some electricity production technologies, as it is the case for example for solar power generation. Similarly to the on-site power consumption, also the calculation of $W_{off}(\mathcal{M})$ depends on the energy $e(\mathcal{M})$ used by the IR model, and the water usage effectiveness $WUE_{off}$ – which in this case refers to the power plant that generates the electricity used by the data center. However, the energy used by the IR model needs to be regulated by the relative amount of energy used by the data center to sustain that power utilization. In other words: for every kWh of electricity used by the components of a server for computation²²2These that can be typically measured are CPU, GPU and memory consumption., additional energy is used by the data center to power e.g., the storage infrastructure, the power units, the pumps and ventilators used in the cooling system, etc.. This additional consumption is captured by the Power Usage Effectiveness coefficient of the data center, $PUE$. As for the on-site water consumption, all these quantities can be parameterized with respect to time. Thus, off-site water consumption is calculated as:

As mentioned, $WUE_{off}(t)$ is dependent on the power plant(s) used to generate electricity – and the electricity used may be produced by a mix of fuels (e.g., nuclear, coal, solar). This is modeled as follows. Be $b_{k}(t)$ the amount of electricity generated using fuel type $k$, and $EWIF_{k}(t)$ be the Energy Water Intensity Factor, measured in L/kWh, for fuel type $k$. Fro example, typical values for coal are $EWIF_{coal}(t)=1.7$, and for nuclear $EWIF_{nuclear}(t)=2.3$ (Li et al., 2023). Then:

Note that typically, and also in the empirical analysis below, values of $k$ and $EWIF_{k}(t)$ are not known to the researchers for small time intervals $t$: instead, they are more likely able to source an estimate of these values based on yearly reporting from their electricity supplier or government authorities.

Table 1 reports the water consumption of common IR models computed according to Equation 1 for experiments performed on the MS MARCO-v1 dataset (Nguyen et al., 2016)³³3See Scells et al. (2022) for the settings of these experiments. Note, we do not re-run their experiments: we use their values for our water consumption models.. Along with water consumption, we also report running time, power consumption and emissions produced – these values are sourced from Scells et al. (2022). To compute water consumption, we used the energy consumption reported by Scells et al. (2022); note however that in Equation 1 $e(\mathcal{M})$ is the Power value in the table divided by the $PUE$ of the data center. Furthermore, we used the annual mean wet-bulb temperature at 3pm in Brisbane (65.3 F), cycles of concentration $S=2.25$, a $PUE$ of 1.89⁴⁴4Which is the $PUE$ of our reference data center, and also that used by Scells et al. (2022)., and a combined off-site water usage effectiveness of $WUE_{off}=1.8$, which is representative of that in Brisbane. These results assume time and season in which models are run do not influence water consumption; we consider the impact of these factors in Section 3.2. The results show an obvious correspondence between power consumption, emissions, and water consumption: as power consumption increases so do emissions and water consumption.

Figure 2 breaks down the water consumption of the considered IR models with respect to on-site and off-site consumption. This figure shows that for each model, under the considering values of the parameters, water consumption is dominated by the need of cooling the data center. This result is highly influenced by (1) the specifications and quality of the cooling tower, which influence the second part of Equation 3, (2) the quality of the water used for cooling, which in turns influences the cycles of concentrations $S$, (3) and the wet-bulb temperature measured at the location of the data center, which in turns is influenced by both the local climate and the season and time of day in which computations occur.

We demonstrate the influence of cycles of concentration, and thus water quality, on on-site water consumption through an example. Recall that the worse the water quality, the more the sediments in the water and thus the need to blow down (i.e. “flush”) the water pipelines of the cooling system of the data center to avoid damages.

For this example we have considered the TILDEv2 model (Zhuang and Zuccon, 2021a) with docTquery expansion (Nogueira and Lin, 2019); other models show similar trends, although this was the IR model with the largest water consumption in the results of Table 1. We have computed the on-site water consumption ($W_{on}$) using the values of cycle of concentrations from the two cooling towers A and B mentioned in Section 3.1, where A was located in a private organization and B in a public hospital ($S_{A}=2.25$, $S_{B}=1.33$): cooling tower $A$ is more water efficient than $B$. We keep all other values the same as those used for Table 1. For site $A$, we obtain $W_{on,A}(TILDEv2)=602.1609$ L, while for site $B$, we obtain $W_{on,B}(TILDEv2)=1261.3524$ L. This example materializes the large impact a lower water quality can have on the amount of water consumed by a data center. The use of high quality water supplies however constitutes an issue per se. First, chemicals could be used to improve water quality: but these add extra costs in the running of the data center operation. These chemicals may also have secondary negative effects on the environment if released into natural water-streams at blow down. Chemicals typically used in cooling towers in fact include corrosion and scale inhibitors, algaecides and biocides, and pH adjusters, which may have harmful impact on the environment (Goni and Mazumder, 2019) – though we note environmental friendly products do exist, e.g., green corrosion inhibitors (Goni and Mazumder, 2019). Second, potable water, i.e. drinking water, is typically of high quality – using this water for cooling means subtracting drinking water to the local population, which is problematic in areas with scarce access to sources of potable water.

Next we discussion the effect of time of day and season on water consumption. This effect is due to the rise of the wet-bulb temperature that is typically associated to the warmer part of the day, e.g., afternoons, or to the warmer seasons, e.g., summer. This can be observed clearly in Equation 3: as the temperature $T_{w}(t)$ raises, so does $WUE_{on}(t)$ and thus consequently $W_{on}$.

To materialize the effect of temperature changes on (on-site) water consumption, in Figure 3 we report the total water usage $W$ of each IR model throughout the year. We base these results on the same settings used to generate the estimations in Table 1, but instead of taking the annual mean wet-bulb 3pm temperature for Brisbane, we consider the monthly mean wet-bulb 3pm temperature. Note that Brisbane is in the Southern Hemisphere, and thus summer is in the period December through to February, and winter is in the period June through August. The figure shows that for the most “thirsty” IR model considered, TILDEv2 with docTquery expansion, water consumption can vary of 192 liters: running this model in the hottest month consumes $23\%$ more water than running it in the coolest month. On the other hand, the impact of season on water consumption for models like BM25 and LambdaMART is minimal (at least in absolute terms).

We perform a similar analysis for showing the impact of time of day. For this, we limit the analysis to the model with highest water consumption, TILDEv2 with docTquery expansion, as an example. We consider two times of the day, 9am and 3pm, and the month with largest difference in mean temperatures at these times in Brisbane: July. The mean 9am wet-bulb temperature in July is 53.6F, and at 3pm is 57.02F: thus the difference is 3.42F. Assuming all other variables are set to the values used for Table 1, the on-site water consumption of TILDEv2 with docTquery expansion at 9am is $W_{on}(TILDEv2,July@9am)=482.90$, while at 3pm is $W_{on}(TILDEv2,July@3pm)=515.1$: a difference of 32.2 liters. Figure 4 extends the comparison between the mean 9am and 3pm on-site water consumption for TILDEv2 with docTquery to all months (for each month, we consider the mean over the last 16 years⁵⁵5Due to the availability of this data from the local meteorology agency.). The figure suggests that for Brisbane the difference in water consumption obtained when running the model at 9am instead than at 3pm is higher during the winter months, than during the summer months. This is because in Brisbane temperature differences between mornings and afternoons are higher in winter than in summer. We note that these findings are location-dependent.