Content uploaded by Lorenzo Posani
Author content
All content in this area was uploaded by Lorenzo Posani on Jun 26, 2019
Content may be subject to copyright.
The carbon footprint of distributed cloud storage
Lorenzo Posania,b,c, Alessio Paccoiab, Marco Moschettinib
aLaboratoire de Physique Statistique, ´
Ecole Normale Sup´erieure, Paris (FR)
bResearch and tech development, Cubbit srl, Bologna (IT)
cto whom correspondence should be addressed, email: lorenzo.posani@cubbit.io
Abstract
The ICT (Information Communication Technologies) ecosystem is estimated to be responsible,
as of today, for 10% of the total worldwide energy demand - equivalent to the combined energy
production of Germany and Japan. Cloud storage, mainly operated through large and densely-
packed data centers, constitutes a non-negligible part of it. However, since the cloud is a fast-
inflating market and the energy-efficiency of data centers is mostly an insensitive issue for the
collectivity, its carbon footprint shows no signs of slowing down. In this paper, we analyze a
novel paradigm for cloud storage (implemented by cubbit.io) [1, 2], in which data are stored
and distributed over a network of p2p-interacting ARM-based single-board devices. We compare
Cubbit’s distributed cloud to the traditional centralized solution in terms of environmental footprint
and energy efficiency. We demonstrate that, compared to the centralized cloud, the distributed
architecture of Cubbit has a carbon footprint reduced of a ∼77% factor for data storage and of a
∼50% factor for data transfers. These results provide an example of how a radical paradigm shift
in a large-reach technology can benefit both the final consumer as well as our society as a whole.
Keywords: Carbon footprint, Cloud Storage, Distributed, Peer-to-peer
1. Introduction
Over the last decades, the general acknowledgment of the climate crisis has driven most western
countries towards an increasing awareness of consumptions and efficiency. As a result, increasingly-
tight policies led to a consistent decrease of the average per-device consumption of household appli-
ances (fridge, cooling, etc.) [3]. However, there is a largely-underestimated factor contributing to
the environmental impact of our daily lives: our use of the Information Communication Technology
(ICT) ecosystem or, in other words, our digital life. An estimation of the total impact of the ICT
ecosystem approaches 1500 TWh of annual consumption [4, 5], which roughly amounts for 10% of
the world energy consumption, more than the energy production of Germany and Japan combined.
A computation of the per-capita consumption shows that the average use of a personal smart-
phone is equivalent, without considering the charging costs, to the energy consumption of an
additional household fridge [5]. In contrast with the trend of the electronics market, however, the
environmental impact of our online life is much less tangible and, as a consequence, much less
controversial. The under-estimation the internet’s footprint, combined with the fast-increasing
trend of online presence and online devices per capita, results in a growing and mostly unopposed
environmental impact that shows no signs of slowing down [6].
Every time a video is streamed from Youtube servers to an iPad, or a photo is accessed on Google
Photos or Dropbox, the whole infrastructure that separates the final user from the corporate data
Preprint submitted to - June 26, 2019
Name Capacity Wattage (peak) PUE Redundancy
Cubbit Cell 1.5 TB 2.55 W 1.0 1.5
HP SO 3620 96 TB 607 W 1.6 2.0
HP SO 5650 2240 TB 6603 W 1.6 2.0
ECS-D5600 2240 TB 9500 W 1.9 2.0
ECS-EX300 192 TB 275 W 1.9 2.0
Storage Pod 480 TB 1500 W 1.6 1.1
Table 1: Equipments - power and capacity of storage equipments
center, as well as the data center itself, has to be powered to reliably transmit information in both
directions. Depending on the relative location of the exchanging nodes, the process of transmitting
information can be orders of magnitude more consuming than storage itself.
In this document we analyze, using an adaptation of the model of Baliga et al. [7], the energy
consumption of cloud storage servicess and compare it to an alternative setup where data is stored
on peer-to-peer low-consumption devices located in users’ houses and implemented by Cubbit.
2. Analysis of centralized cloud consumptions
The energy consumption of a cloud storage service can be divided into two main factors:
1. the cost of storing the data, i.e. powering and cooling the data center (Storage consumption)
2. the cost of sending the data from the user to the server and back (Transfer consumption)
While the first can be estimated from technical specifications of storage equipment, the second
needs a more detailed analysis that takes into account the public internet infrastructure and the
geographical distance between the user and the server. For both these estimations we refer to the
model of Baliga et al. [7], where energy consumption is computed accounting for several factors,
including the multiplicity of involved devices, redundancy, cooling, overbooking (see below).
To delineate the calculation, we start from the storage consumption, i.e. the average power,
expressed in W/TB, necessary to store the payload in hot storage. We updated the technical
specifications with respect to [7], as hard-disk storage capacity has dramatically improved in the
last years. As a model for data center racks, we considered five of the most recent products from
three leading companies in the sector of enterprise storage hardware, see Tab.1. For each storage
appliance we take the peak consumption, as it is the information made available in the manufacturer
specs sheet [8, 9, 10]. To estimate the capacity, we consider every rack to be filled with 8TB disks
(estimated from the mean of HDD dimensions in a recent BackBlaze report [11], ≃7.2 TB/Disk)
that are fully employed to store customers’ cloud files (with no empty space overhead). When
the heat produced is not specified in the manufacturer specs sheet (as in the case of HP racks)
we consider a 1.6×cooling factor, estimated by the most recent reports on average Power Usage
Effectiveness (PUE) in data centers []. Similarly, when the redundancy strategy is not explicitly
stated, we use a 2×factor [7] as if every files was mirrored two times in the same or in a different
data center. Both BackBlaze and Cubbit employ a redundancy protocol based on erasure coding
with replication factors 1.1×and 1.5×, respectively. These values are simply combined in the
2
Equipment Capacity Consumption
Data Center gateway router Juniper MX-960 660 Gb/s 5.1 kW
Ethernet Switch Cisco 6509 160 Gb/s 3.8 kW
BNG Juniper E320 60 Gb/s 3.3 kW
Provider Edge Cisco 12816 160 Gb/s 4.21 kW
Core router Cisco CRS-1 640 Gb/s 10.9 kW
WDM (800 km) Fujitsu 7700 40 Gb/s 136 W/channel
Table 2: Equipments - power and capacity of routing equipments. Data from [7]
following formula to obtain the storage consumption per Terabyte:
Pstorage
dcenter = PUE ×redundancy ×peak Wattage
n disks ×8 TB (1)
Likewise, we compute the transfer energy, expressed in J/GB, following the public internet
model of [7]. The analysis relies on the definition of the consumption per bit, which is computed
by dividing the operating power (W) by the total transfer capacity (Gb/s), resulting in a Joule/bit
measure, then converted in J/GB. These units are taken from the manufacturer specs sheet, shown
in table 1. These quantities are then combined with a set of coefficients reflecting the redundancy
of the packet transmission, the under-operating regime of the infrastructure and the cooling energy,
and the multiplicity of some devices in a single transmission (e.g. two ethernet switches at entry
points plus another one inside the data center). The average distance between core routers on the
network is estimated to be c.a. 800Km. For a full description of coefficients and estimations we
refer the reader to [7].
Etransfer
dcenter = (2)
= 6×3Pes
Ces
+Pbg
Cbg
+Pg
Cg
+ 2 Ppe
Cpe
+ 18 Pc
Cc
+ 4 Pw
Cw
≃23.9kJ
GB ,
where the prefactor of 6 accounts for redundancy (×2), cooling and other overheads (×1.5),
and the fact that todays network typically operate at under 50% utilization (×2); the addends
represent, in order, the ethernet switch, the broadband gateway, the data center gateway, the
provider edge router, the core network, and the relay optical fiber transmission. The detailed
analysis of pre-factors can be found in [7]. Briefly, the factor 3 in the ethernet switch accounts for
the two routers involved in the access to the public internet plus the router located inside the data
center; the factor 18 in the core network accounts for an average of 9 hops (2 baseline + 7 for the
800km distance between core nodes) of internet packets from source to destination, times 2 for the
redundancy.
3. Analysis of Distributed Cloud consumptions
The distributed architecture of the Cubbit network relies on the same public internet infras-
tructure delineated in the previous chapter. However, the distributed paradigm has three key
3
differences from server-based cloud storage:
1. The low energy consumption of storage devices (based on the Marvell ESPRESSObin [12])
2. The absence of cooling overhead
3. The geographical proximity between the users and their stored data
We consider a network of Cubbit Cells [13], each composed by an ARM-based SBC and a HDD
(Western Digital Blue) of 1 TB or 2 TB (estimated average 1.5 TB disk). Each Cell is located in
a user’s house and connected to internet by an internet service provider. Files on Cubbit’s cloud
are stored with a redundancy factor of 1.5 (Reed Solomon erasure coding with 24 + 12 redundancy
shards [1, 14]). As done for the centralized cloud, we analyze the consumption of both storage
(W/GB) and transfer (J/GB).
The Marvell ESPRESSObin has a single-core peak consumption of ∼1W [12], while the em-
bedded WD Blue HDD has a peak consumption of 1.4 W (1 TB) and 1.7 W (2 TB). We here
assume that half of the network is composed by 1 TB devices and half by 2 TB devices, giving an
average storage of 1.5 TB, corresponding to an average peak consumption of 1.55W. The storage
energy consumption of the Cubbit network is therefore computed as
Pstorage
cubbit = 1.5×1 + 1.55 W
1.5 TB ≃2.55 W
TB .(3)
In Cubbit, shards of the distributed payloads are preferably distributed in Cubbit Cells that
are located in geographical proximity of the user, since the distribution of the shards is controlled
by the AI optimization routines of a coordinator server[1]. We therefore consider the scenario
where data is stored in nodes at an average distance of 80 km from the user’s access point. In this
scenario, we can assume an average number of 2 packet hops in core network routers. This lowers
the corresponding factor 18 in Eq. 4 to a factor 4, accounting for two core hops and the redundancy
of the packets on the network (factor 2). For the same reason, the 800km-relay consumption Pw
is not taken into account. With respect to Eq. 4 we also ignore all data-center specific terms: one
ethernet switch and the data center gateway. However, we need to consider an additional BNG,
since transfers are performed through p2p connections between endpoints located within an ISP
network. The transfer energy per GB is therefore computed as
Etransfer
cubbit = 6 ×2Pes
Ces
+ 2 Pbg
Cbg
+ 2 Ppe
Cpe
+ 4 Pc
Cc(4)
≃11.9kJ
GB .
4. Comparison between centralized cloud and Cubbit distributed cloud
As a first analysis, we show in Fig. 1 the comparison between Cubbit and the centralized cloud
in terms of pure storage consumption. Cubbit achieves a reduction ranging between ∼50% and
∼95% with respect to data centers’ racks. By taking the average over racks as a mid-range
estimation (although we have no information on the relative distribution of these racks in the data
center market), we obtain
∆Pstorage =Pstorage
dcenter −Pstorage
cubbit ≃9W
TB ,(5)
4
corresponding to an overall 77% reduction:
∆Pstorage
Pstorage
dcenter =≃0.77 .(6)
Figure 1: Comparison between centralized and distributed clouds in terms of peak consumption per TB of stored
data.
Similarly, the difference in terms of transfer energy per GB is
∆Etransfer =Etransfer
dcenter −Etransfer
cubbit (7)
≃12.0kJ
GB = 3.33 kWh
TB ,
which corresponds to a 50% reduction of the energy needed to transfer data from the cloud to
the user, and back.
The reduction of carbon footprint of Cubbit compared to centralized solutions can be computed
by comparing the storage power and the transfer energy for typical use case, such as backup plans
and frequent access of, for instance, a web-hosted video.
4.1. Backup
A backup service hosted on the cloud is characterized by large volumes that are not frequently
accessed. In the context of the carbon footprint, the consumption of a backup plan will, therefore,
5
Figure 2: Comparison between centralized (DCenter) and distributed (Cubbit) clouds in terms of annual carbon
emissions per TB of stored data.
be dominated by the storage term. If we consider a storage plan for a professional backup of 25
TB with very small daily access, we find that the total energy saved in a year is
∆E(25 TB backup) = (8)
25 TB ×∆Pstorage ×365 ×24h≃1971 kWh .
By considering a rough factor of 0.5 KgCO2 for each kWh of consumed energy[15], the use of a
distributed cloud over a centralized one would correspond, for such a backup plan, to a reduced
carbon emission of c.a. -1000.0 kgCO2/year. If we consider that data centers usually operate on
the Petabyte scale, we easily see that the yearly reduction in emissions scales up to hundreds of
tons of CO2
∆Footprint(backup) ≃40 000 kgCO2/year/PB (9)
4.2. Streaming
The reduction in consumed energy and, consequently, in carbon emission is significantly larger
when considering large volumes of data transfers. For example, if we consider a medium news
website hosting 25 TB of data and streaming, on average, 10 TB of data per day (e.g. 10,000
visualizations of 100 MB each) the saved energy per year would be
∆E(10 TB streaming) =
25 TB ×∆Pstorage ×365 ×24h
+ 365 ×∆Etransfer ×10 TB
= 14136 kWh ,(10)
6
Figure 3: Comparison between centralized (DCenter) and distributed (Cubbit) clouds in terms of annual carbon
emissions per TB of daily streamed data.
which roughly corresponds to -7,000 kgCO2 emitted per year. Note that these computations
assume that data are broadcasted to a local audience. While this might be the case for university
data, local news, or targeted marketing, it has a limited range of applicability that has to be taken
into account.
4.3. Large scale: global cloud industry
Finally, if we speculate about the overall data volume of a global consumer cloud storage service
like, for example, Dropbox or Google, values rise dramatically. Such interpolations have to be taken
with due caution, since estimations are based on undisclosed values. For the sake of speculation, we
consider a use base of c.a. 600 millions users. The latest disclosed conversion rate from free (2 GB)
to premium (2 TB) is around 3%. This results in a theoretical data volume of ca. 37.2·106TB of
storage. Considering a factor 5 due to overbooking, it gives an estimation of 7.4 106TB of effective
cloud storage. We can make a conservative estimation that each user transfers, on average, 50 MB
of files from/to the cloud, which implies a daily transfer volume of c.a. 190 TB. If we plug these
estimations in our model, we obtain a total saved annual energy, using a distributed architecture
rather than a centralized one, of ∼6.7·108kWh, equivalent to saving carbon emissions in the
order of 300 million kgCO2 per year.
Acknowledgements
The authors are grateful to S. Baldi, A. Albani, and A. Rovai for valuable comments on the
manuscript and fruitful discussion.
7
[1] L. Posani, M. Moschettini, A. Paccoia, Cubbit: a distributed, crowd-sourced cloud storage service, Invited Talk
at CERN Workshop on Cloud Storage Services (CS3 2018) (2018).
[2] L. Posani, M. Moschettini, A. Paccoia, Cubbit, the distributed cloud, Invited Talk at CERN, CNR Workshop
on Cloud Storage Services (CS3 2019) (2019).
[3] Energy efficiency and energy consumption in the household sector, https://www.eea.europa.eu/
data-and-maps/indicators/energy-efficiency- and-energy-consumption-5/assessment, web page.
[4] G. International, How Clean is Your Cloud? Catalysing an energy revolution, Technical Report, 2012.
[5] M. P. Mills, The cloud begins with coal, Digital Power Group (2013).
[6] Cisco vni global fixed and mobile internet traffic forecasts, https://www.cisco.com/c/en/us/solutions/
service-provider/visual-networking-index- vni/index.html, web page.
[7] J. Baliga, R. W. Ayre, K. Hinton, R. S. Tucker, Green cloud computing: Balancing energy in processing,
storage, and transport, Proceedings of the IEEE 99 (2011) 149–167.
[8] Manufacturer specs sheet, https://www.hpe.com, web page.
[9] Manufacturer specs sheet, https://www.dellemc.com, web page.
[10] Manufacturer specs sheet, https://www.backblaze.com, web page.
[11] Blackblaze report, https://www.backblaze.com/blog/2018-hard-drive-failure- rates/, web page.
[12] Marvell espressobin specs sheet, https://www.http://espressobin.net/, web page.
[13] Cubbit website, https://www.cubbit.io/technology, web page.
[14] M. Monti, S. Rasmussen, M. Moschettini, L. Posani, An alternative information plan, Technical Report, Working
paper Santa Fe Institute, 2017.
[15] Carbon trust energy and carbon conversions,
8
