IEEE Consumer Electronics Magazine - May/June 2023 - 36

DNN Migration in IoTs
FIGURE 5. Comparison of various DNN partition-uploading methods. (a) Alexnet model. (b) VGG model.
When compared with the Direct Uploading algorithm
that does not adopt the DNN partitionuploading
plan, the IONN algorithm and Recursive
Efficiencyalgorithm using theDNNpartition-uploading
plan have a lower total system delay to complete
the DNN model query. This is mainly due to
the DNN partition P1 being too large and the transmission
delay of uploadingthewholeDNNpartition
P1 being too high. In the process of uploading the
DNN partition P1, DNN queries will only be completed
by local devices, and the insufficient computing
capacity of local devices leads to the high
delay of DNNqueries. In addition, the RecursiveEfficiency
algorithm has better query results than that
of IONN algorithm, which is mainly obtained by
more fine-grained partitions P3.In summary, we
know that adopting the DNN partition-uploading
plan can obtain more efficient DNN query performance,
and more fine-grained DNN partition P3
helps to obtainmore efficient DNN query results. In
short, the abovementioned experimental results
demonstrate that the method proposed in this section
can extend different types of DNN partitionuploading
plans to a multiserver environment. Furthermore,
the comparison with the direct uploading
algorithm also illustrates the necessity of extending
the DNN partition-uploading plan to a multiserver
environment.
CURRENT CHALLENGES AND
FUTURE DIRECTIONS OF DNN
MIGRATION
Along with the increasing number of IoT devices,
people have new demands for service quality
36
in daily life. In addition to the large granularity of
DNN partitions and not being suitable for multiserver
environments, DNN migration also faces
some challenges and opportunities in real-world
scenarios.
On the one hand, applications such as
unmanned driving, virtual reality (VR), and augmented
reality (AR) all put forward strict requirements
for the delay. However, processing massive
amounts of data on the IoT devices puts a heavy
burden on battery consumption, and the high data
transmission delay of the cloud computing center
also brings greater transmission energy consumption
pressure on the client. Furthermore, The
requirements of users for QoS are constantly
increasing, and they hope to obtain the results of
DNN migration at a lower cost. In summary, when
DNN migration is performed in a resource-limited
environment, how to achieve the overall system
optimization of latency, energy consumption, and
cost andhowto specifically adjust the optimization
degree of each optimization objective according to
user requirements is an urgent issue tobe solved.
On the other hand, we consider practical
issues that are ignored in the process of DNN
model migration. These issues are caused by the
greedy occupation of resources by users and the
unbalanced metric optimization. It is known that
the users entering the computing environment
are constantly changing, so when the users
entering the environment first occupy all the
computing resources of some edge/cloud servers,
the users entering the environment later will
not be able to use the above servers, resulting in
IEEE Consumer Electronics Magazine

IEEE Consumer Electronics Magazine - May/June 2023

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - May/June 2023