The energy consumption of IT equipment has recently become a major concern. In particular, traditional datacenters have made the news headlines, as a result of calculations showing that they accounted for over 1.5% of US energy consumption in 2006. This share is expected to increase significantly, in view of traffic forecasts. This lead some authors to predict that the carbon dioxide emissions associated with such datacenters will surpass those of the airline industry by 2020.

Due to the current threat of global warming, the European Commission has issued guidelines on the energy efficiency of IT equipment. Best practices for energy management in datacenters have been discussed in the press, notably with Google explaining how they achieve a low “PUE” (Power Usage Efficiency) that is the ratio of total energy expenditure against the sole consumption of IT equipment.

Beyond improving the efficiency of the individual components within the IT equipment and minimising the associated energy costs for cooling, power transmission and conversion, alternative architectures can be applicable on this issue. Bearing this in mind, the aim of this deliverable is to conduct an evaluation of how the NANODATACENTERS architecture could potentially change the overall energy cost of datacenter services, focusing particularly on Video-on-Demand applications.

To perform such an analysis, measurements of the power consumption of commercial Video-on-Demand servers were made and compared with the power consumption of current Triple-play gateways used to serve video. A key measurement result for both types of devices was that over 80% of their peak power consumption was spent by simply waking them from the standby/ idle state. On top of this fixed, baseline power consumption, the additional power depends linearly on the speed at which video is served from the devices with the slope of this linear dependency significantly smaller for gateways than for servers.

These basic energy consumption profiles can be used to compare the energy cost of serving videos from servers versus gateways. If the baseline power consumption of gateways is discounted, the energy cost of serving content from gateways is as little as 50% of the corresponding energy cost when relying on servers.

Discounting this baseline energy consumption on gateways can be motivated as follows: if content is served only from a gateway that is already in use for another purpose, then the baseline consumption would be present no matter whether the content is served from that gateway or a server. Furthermore, current Triple-play gateways remain “always on”, if only to offer uninterrupted telephony services. Other aspects taken into account were the PUE of datacenters, the transmission and conversion costs incurred in the power grid and the energy consumption of the routers in the network.

It should be noted that the Video-on-Demand workload that can be supported by the NANODATACENTERS approach depends on several complex parameters, such as the peak uplink bandwidth and the storage space per gateway. The type of workload also affects the overall efficiency: “skewed” workloads (i.e. where demand is dominated by a few popular applications) are more easily cacheable, and result in a larger efficiency in the NANODATACENTERS approach. Conversely, large catalogues of content are harder to cache, thus reducing the efficiency of the NANODATACENTERS approach as the catalogue size increases.

In order to provide an in-depth analysis of the energy efficiency of the NANODATACENTERS architecture with all the above stated factors, an extensive event-driven simulation was developed and run to assess the impacts of each factor in isolation. These imulations were driven by three distinct representative workloads: (i) an IPTV trace collected from Telefonica users, (ii) the Netflix trace provided in the Netflix prize contest, and (iii) a YouTube trace. All three correspond to real-life distinct types of content catalogues, any of which could potentially be provided by a NANODATACENTERS platform.

Two particular scenarios of interest were identified. In the first scenario, all gateways are “always on”, while in the second scenario, gateways are assumed to be only “on” when the corresponding home is the originator of a request for content. Clearly, the latter scenario will lead to the least amount of energy savings for the NANODATACENTERS approach. In summary, the simulation findings show that the NANODATACENTERS approach is capable of serving almost the entire workloads which were considered, whilst achieving a 40% decrease in its energy consumption with the “always on” gateway scenario and having only 10GB of storage per
gateway. However, in the second scenario, more storage has to be installed on each gateway in order for it to serve a significant fraction of the workload. In this case, energy savings of 28% can be typically made with 128GB of storage per gateway. For an extensive set of results and related discussion, please refer to the full paper.

Details can be found in the report.