IEEE Circuits and Systems Magazine - Q1 2023 - 64

provided in Tables 1 and 2 are derived from values reported
in the literature. For example, the constant offset
power ps,UT,d,r,0 = 805 mW in the left column of Table
1 corresponds to the idle power in airplane mode reported
in [43]. To obtain the receiving power, we then
subtracted this idle power from the reported overall
video playback power as ps,UT,d,r,Rx = 1084 mW − 805mW
= 279 mW (for the low-end video power). The other values
were derived analogously.
The values in Table 1 show that the power consumption
depends significantly on the device. Furthermore,
various parameters are shown in a certain range, as
reported in the referenced studies [16], [43], [57]. The
variations can be attributed to the video parameter settings
(bitrate, frame rate, etc.) and other configuration
options.
For more accurate modeling of the overall end-user
energy consumption, one would have to consider the
dependency of the powers, i.e., ps,UT,d,r,0, ps,UT,d,r,Rx, and
ps,UT,d,r,Tx, on additional parameters such as video properties
and screen brightness settings. This relation was
studied extensively in the literature [16], [42], [45], [46],
[47], [48], [49], [50], [51], [52], [53] and could be included
in the model proposed here. However, in this article,
we rather focus on a realistic range of power consumption
values, which is sufficient for modeling the impact
of high-level design choices of the system engineer in
Section IV. A more detailed modeling approach could be
developed in future research.
C. Video Providers
To model the yearly provider-side energy consumption
of a single service, we can choose either a bottom-up
or a topdown approach. For the top-down approach,
one could use statistical information of the
yearly energy consumption of all worldwide data centers
(DCs) and their corresponding share of workload
for online video systems [9], [59]. For general-purpose
DCs in the cloud, a large amount of research has been
performed, which analyzes the overall power consumption
in detail [60], [61], [62]. As a drawback, the
energy values obtained in this manner provide no information
on the power consumption related to the
specific devices and the exact algorithms that were
deployed in the DC.
Hence, in this article, we follow the bottom-up approach
and consider single devices in DCs as well as
CDNs, which are servers. Hence, we split up the energy
consumption of DCs into the energy consumption
of single servers. For DCs explicitly targeting online
video applications, the literature on energy modeling
focuses on CDNs [17], [39], [40], [63], which we take as
the basis for our considerations. As a consequence,
64
IEEE CIRCUITS AND SYSTEMS MAGAZINE
Figure 2. Tasks of a provider to operate an online video
service.
we model the yearly energy consumption of a CDN
as the sum of the energy consumption of all related
servers as
EEVP σ,,, ,
σ∈Σ
ss
s
VP = ∑ησ
(6)
where σ is the index for a single server, ∑s is the set of
servers used by the s-th service, ησ is the PUE of the
σ-th server, and Es,VP,σ denotes the yearly energy consumption
of the σ-th server.
Based on the modeling of a CDN presented in [17],
the tasks of the servers are defined as follows (see
Fig. 2): First, the server receives input videos (a movie,
a live stream from, e.g., a football match, a video from
a user to share via social networks, or a live teleconferencing
stream). Depending on the type of service,
it is then transcoded to different codecs or bitrates to
generate additional videos. Afterward, in the case of
on-demand video, the videos are stored on one server
(e.g., in a DC) or on multiple servers in a CDN. In the
case of a CDN, the video is transmitted to a set of surrogate
servers, which store copies of the video to reduce
the spatial distance to the end user [17]. Finally, the
videos are transmitted to the end user in a single-cast
or multi-cast fashion triggered by user requests. Note
that depending on the video streaming application,
some of these tasks may not be relevant. For example,
the provider of a peer-to-peer (P2P) service (e.g., video
conferencing) does not have to receive, transcode, and
FIRST QUARTER 2023
IEEE Circuits and Systems Magazine - Q1 2023

Table of Contents for the Digital Edition of IEEE Circuits and Systems Magazine - Q1 2023
