1.
Edge Raspberry Pi (The first layer)
The first layer is the bottom level called edge raspberry pi (ERP), whose main task is the
response to and implementation of edge tasks, such as smart home, object detection, and face
recognition.
2. Monitoring and scheduling raspberry pi (MSRP) ( The middle layer)
The monitoring of RP on edge nodes at the bottom level.
transmit the tasks they cannot process to CC by the allocation scheduling task strategy.
(hình 3)
3. CC (Cloud Center)
CC is at the top level
When CC receives the tasks edge nodes cannot process, it processes these tasks.
The design objectives of the system mainly include edge–cloud collaboration flexibility, time
delay minimum ,priority in the processing of edge tasks.
Equation Meaning Result
Processing time at the The ERP executes the task locally.
edge node (ERP) The time depends on the computational
(1) workload cv and the processing rate pi .
cv : computation If the ERP has enough resources, it performs
volume the task locally to reduce delay.
pi : ERP processing
speed
(2) Offloading ERP cannot process (Iv,i=0)
preparation time estimates the time required to send those tasks
(when ERP is to the upper layers (MSRP or Cloud).
overloaded)
(3) Data transmission Measures how long it takes to transfer data from ERP
time → MSRP → Cloud.
Depends on data size (d) and network
bandwidth (ri).
(4) Task scheduling time The MSRP spends time managing and deciding which
tasks should stay at the edge and which should be
offloaded to the cloud.
(5) Cloud processing Heavy tasks that cannot be processed at the
time (CC) ERP are sent to the Cloud.
Cloud servers higher processing power
this time is usually shorter
*Objective function for total delay optimization
TERPi: processing time at the edge node
Tschedi,j: coordination time between nodes
Ttransi,j: time to transmit data to the cloud
Tcloudi: cloud processing time
allocate tasks appropriately (which tasks are kept in ERP, which tasks are sent to Cloud)
the total delay time is minimized.
*Edge Optimal Power Search Algorithm
M= minimum transmission power
N= maximum transmission power
ε=10−4: allowable error
while loop (N - M > ε): iterate until the search interval length is less than the allowable error
repeat the process until the error is small enough.
*Analysis of platform performance of new edge - cloud collaboration
To evaluate the effectiveness of the Edge - Cloud collaboration system, a research
implementation 6 typical computing tasks, including:
Health management
Smart home
Object recognition (object detection)
Face recognition
Smart factory
License plate recognition
This task is compared in three modes of operation: Cloud only, Edge only, Edge–Cloud results
(collaboration)
Results
When the edge node RP performs face recognition for the first time, the model loading
time is 2.14 seconds, and the results are displayed relatively quickly.
However, when another edge node (RI) carries out an object detection task, the system
needs to load the SSD (Single Shot Multibox Detector) training model for target
detection.
→ This causes a longer delay for detecting the first face.
The Edge–Cloud collaboration mode achieves a faster response time compared to using
Edge-only or Cloud-only modes.
The frame rate of the Edge–Cloud mode is almost the same as Cloud-only, and
significantly higher than Edge-only.
The system operates more smoothly, especially in real-time recognition tasks such as face
and object detection.
The combination of Edge and Cloud offers the following advantages:
Significantly reduces response time.
Maintains a high frame rate, supports real-time data processing more effectively than
using Edge or Cloud alone.
Comparison of platform performance on different modes. (a shows the comparison of task time delay, and b demonstrates
the comparison of framework rate).
* Analysis of simulation results of optimal edge - cloud unloading strategy
Figure 10 presents a comparison of task execution performance among different offloading
strategies.
TCO (Task Caching and Offloading): an algorithm based on energy efficiency,
combining task caching and offloading to the system.
CLA (Caching and Local Execution Algorithm): performs local caching and task
execution directly on the edge device.
CEA (Caching and Edge–Cloud Execution Algorithm): integrates local caching with
joint edge–cloud task execution.
Results
As the number of users increases, the total energy consumption of the system
gradually decreases.
→ This happens because, with more devices available, users have more options to
choose where to execute tasks (Edge or Cloud), leading to more efficient resource
utilization.
In comparison:
o The proposed Optimal Matching and Fast Matching algorithms achieve faster
and more energy-efficient task completion,
o While still satisfying time delay constraints required by the system.
Compared with other algorithms, the optimal edge–cloud offloading strategy can:
Reduce the total energy consumption of the system,
Improve task execution efficiency,
And maintain low response time — making it suitable for large-scale IoT
applications.
Conclusion and Future Work
In the context of the Internet of Things (IoT), meeting the requirements of delay-sensitive
applications is challenging due to the long physical distance between the cloud center (CC)
and the sensors. This distance often makes it difficult to guarantee high-quality service (QoS).
To overcome this limitation, this study combines cloud computing and edge computing to
design and implement a new edge–cloud collaboration framework.
Limitations and Future Research
Due to time constraints, the research still has several limitations:
The current framework only considers remote command execution, data streaming
transmission, and file transmission.
Lightweight node tasks have not yet been virtualized or containerized.
In future studies, the authors plan to:
Implement virtualized lightweight edge tasks,
Integrate with Kubernetes (K8s) and lightweight container technologies (e.g., Docker)
to improve scalability and flexibility.
This diagram illustrates a Multi-Tier Wireless Sensor Network (WSN) Architecture
integrated with Cloud Computing.
1. Bottom Layer (Sensing): Sensor nodes in the monitored area collect data.
the data is aggregated and sent to the sink node.
2. Mid-Layer (Aggregation & Processing): Data flows from the Sink node and
local groups (Node 1, 2, 3...) up to scheduling nodes. These nodes perform
preliminary processing and resource management. The mission management
center also handles local data and external network communication (Internet,
satellite).
3. Top Layer (Cloud Computing): The Scheduling nodes forward the processed
data to the Cloud Center (via wireless link dcv). The Cloud provides vast
computing power for deep analysis, large-scale storage, and final service
delivery.