Content for  TR 22.882  Word version:  19.3.0

Global warming caused by excessive emissions of GHG (e.g., carbon dioxide) due to human activity (e.g., burning fossil fuels for electricity generation) is the main driver to climate change, which poses a significant threat to society and the environment. To achieve carbon neutrality, it is important to reduce the GHG incl. carbon emissions in the first place rather than offset them later. Recent advancements in communication and computing capabilities of networks (e.g., 5GS, cloud services) enables offloading tasks to networked and distributed computing nodes (e.g., edge computing, cloud computing) for a wide range of services. However, the rising demand for such services in turn triggers a rising demand for energy and a greater risk of an even higher resulting GHG footprint. 3GPP plays a crucial role in the ICT sector to enable the deployment of these technologies on a global scale and therefore must also play a central role in enabling a sustainable future. The adoption of alternative sustainable sources of energy incl. renewable energy (e.g., solar, wind, hydropower, geothermy) could help offset the GHG footprint of energy generation based on fossil fuels, even though their corresponding environmental impact also need to be considered. From an ICT standpoint and, 3GPP system in particular, the energy used by computing nodes in networks can be from varied energy with different related levels of environmental impact incl. GHG emissions. Due to the highly variable and unpredictable nature of renewable energy sources, the supply of renewable energy varies substantially by time and location. Hence, it is critical to take temporal and spatial dimensions of energy sources into account to accomplish compute tasks not only for a better traceability of the energy sources used but in turn for enabling a more sustainable energy use to achieve those tasks. Up until now usually a system is designed to finish compute tasks as soon as possible (high throughput) and indicate results to the requester as soon as possible (low latency). However, some compute tasks have flexibility in both when and where they are executed, i.e., such type of workload could be executed in any computing node and tolerate some delays if the workload gets completed within certain given deadline. For example, some of AI/ML training, simulation, and video processing tasks might not require a quick response, which would allow flexibility to delay the execution of the related workloads in a computing node until, e.g., the utilized energy is deemed satisfactory in terms of GHG emissions. Such flexibility further allows to route workloads to a computing node using the (most) sustainable energy sources at that moment. As part of service, 3GPP system is able to execute compute tasks in a sustainable way by leveraging such flexibility. In addition, consuming the renewable energy immediately when they are available, instead of storing them for the future use (e.g., in a big battery system), can also bring some economic benefits to operators or service providers, because this can reduce the cost and investment for scaling the energy storage system needed by the overall system. In the following use case, by considering the temporal and spatial information of sustainable energy source and availability, the possibility of reduction of the GHG footprint for application services is explored.

The operator A provides the computing services through the computing nodes owned by itself or other third-party companies via certain Service Level Agreement (SLA), which execute the compute tasks (e.g., offloaded by users). Each computing node is powered by renewable energy (e.g., solar energy), non-renewable energy (e.g., coal) or both. The highly variable nature of renewable energy sources makes the resulting GHG emissions by each computing node varies considerably by time and location. The high cost of large-scale energy storage system (e.g., battery system) also brings the incentive to the operator to consume the renewable energy immediately when it is produced (e.g., to reduce the cost for building the needed battery system). The ratio of renewable energy measures the ratio of the power that is used from renewable energy sources as a percentage of total power usage in a given time unit. Eva is an AI engineer who needs to train some AI/ML models for her research work. Eva has collected all the needed data (e.g., the images of cats and dogs) during the weekdays. To train this model, the required dataset must be sent to a computing node, and the node will train the specified model (e.g., a dog/cat classifier) over this dataset. Eva needs to get the training result at the beginning of workday next week. Her compute tasks for AI/ML model training are offloaded to the system owned by the operator A for execution. The operator A offers a "green compute and communication service" which can decide when and where the offloaded tasks are computed to reduce the overall GHG footprint of the system. This green compute and communication service requires tolerated deadline of compute task specified by the user, i.e., the quality of experience is not degraded as long as the compute task is finished within the given deadline. Eva loves our planet, so she is using this service for reducing the GHG footprint of her research work.

1. Eva subscribes the green compute and communication service to save our planet.
2. Eva indicates to the operator A that the compute task needs to be finished before the next workday (8:00 AM on Monday).
3. Eva offloads the compute task of AI/ML model training to the system owned by the operator A before she left the office (7:00 PM on Friday) in New York.
4. In the operator A's system, the "computing node NY" (i.e., the computing farm located in New York) is the closest computing resource to the Eva's workplace. Traditionally the "computing node NY" is selected to execute Eva's task immediately; however, there is no solar power in New York at this moment (i.e., the ratio of renewable energy is low).
5. If Eva's AI/ML model is trained by the "computing node NY", it will result in some GHG emissions to the air which is not friendly to the environment.
6. Fortunately, the "green compute and communication service" has two alternative options for the execution of Eva's compute task based on the ratio of renewable energy reported by the "computing node NY" and another node "computing node LA" located in Los Angeles:
   • [Option 1: Greener Location] The "computing node LA" located in Los Angeles (is on 4:00 PM) having abundant solar energy at that moment (i.e., the ratio of renewable energy is high). The dataset can be sent to "computing node LA" and the results are sent back to Eva after the completion. Since the execution will not last over one day, the system can adopt this option even if it requires more time for the communications.
   • [Option 2: Greener Time] The "computing node NY" will have plentiful solar energy during the period of 9:00 AM - 4:00 PM every day. The training executed during the daytime of the weekend will not generate any GHG emissions. Since the task can be finished before the next workday, the system can adopt this option to schedule the training to be executed during the weekend.
   In addition, by consuming the renewable energy immediately when it is produced, the operator can reduce the scale of its renewable energy storage system and reduce the overall cost.
7. By adopting the either option provided by "green compute and communication" service, the execution of AI model training requested by Eva can be nearly carbon-free and Eva still obtains the desired training result before the deadline.

Eva's AI/ML model training is finished before the targeted deadline while protecting our beautiful planet. Operator A reduces the scale of its renewable energy storage system and reduce the overall cost.

None.

[PR.5.14.6-1] [PR.5.14.6-2]