Page 241 - Cloud computing: From paradigm to operation

Framework and requirements for cloud computing 1

– Edge cloud: The edge cloud is deployed at the edge of the network accessed by CSCs and has a small
resource capacity. The edge cloud requires specialized hardware resources on purpose; i.e., the
resources in the edge cloud are constrained due to limitations of space or power. The edge cloud
may have different configurations of resources and cloud capabilities types with physical and virtual
resources depending on a CSC's requirements of cloud services and conditions in the deployment
environment.
Figure 6-2 shows an example of providing real-time service on distributed cloud for the case of a machine
learning (ML) service. The example has four phases.
– Phase 1 – collecting: End device 1 transfers sensor data to an endpoint of ML training service in the
core cloud to train the data;
– Phase 2 – training: ML training service trains the data and gets a trained rule at the core cloud;
– Phase 3 – caching: The trained rule is cached from core cloud to regional cloud; the trained rule is
then cached to edge clouds 1 and 2;
– Phase 4 – forecasting: End devices 1 and 2 transfer sensor data to edge clouds 1 and 2, and then
edge clouds 1 and 2 perform ML forecasting in real-time processing. The forecasting result is quickly
delivered to end devices 1 and 2.

If the ML forecasting service is also processed at the core cloud without the edge cloud, then the end device
of the CSC will receive the forecasting results from the core cloud with high latency. The example in Figure 6-
2 highlights how the distributed cloud provides benefit to the cloud service in the edge cloud by providing
low latency and real-time processing.

Figure 6-2 – Example of providing real-time service on distributed cloud
for the case of a machine learning service

NOTE 3 – In this example, it is assumed that the network latency from the end devices to the edge cloud is much lower
than from the end devices to the core cloud, and that the execution time of the ML forecasting service is faster than the
ML training service.

233