IEEE Consumer Electronics Magazine - May/June 2023 - 29

of three parts: DNN model preprocessing, partition-offloading plan, and partitionuploading
plan. Accordingly, we introduce the operation of the DNN migration and the
available methods for each part. In addition, we improve the DNN partition-uploading plan
in a multiuser edge-cloud collaborative computing environment. Finally, we highlight the
important challenges of achieving more efficient DNN migration and point out the
unresolved issues of DNN migration, which may shed light on future research directions.
& IN RECENT YEARS, the Internet of Things
(IoT) and mobile Internet have developed rapidly,
smart cities, smart homes, and smart transportation
have become an indispensable part of social
life. As the core of artificial intelligence, deep learning
technology, especially deep neural networks
(DNNs), has been widely used in a variety of fields,
which also brings many new challenges to mobile
terminals with limited resources. The requirements
of emerging intelligent applications on communication
systems are mainly reflected in four
aspects: low response time, low energy consumption,
low monetary cost, and balanced-computing
resource allocation. The most obvious contradiction
is between the limited computing capacity of
IoT devices and running complicated DNN inference.
When DNN models implement inference in
a resource-constrained computing environment,
how to optimize objectives, such as delay, energy,
andmonetary cost, is of great significance.
Considering that the limited computing resources
of mobile devices cannot support complex
DNN operations, the traditional approach is to offload
partial DNNmodels to remote cloud servers to
reduce the pressure of local devices. DNN migration
in edge, cloud, or fog computing environment,
as well as other computing platforms1 have
attracted great attention. As we all know, cloud
computing has a high computing capacity and sufficient
storage space. However, the cloud server is
usually far away from the local device, and the data
transmission between the cloud server and the client
is readily affected by various factors, such as
transmission delay, data volume, and central computing
capacity. These factors put great pressure
on the network bandwidth and easily cause excessive
transmission delay, which cannot meet user
quality of service (QoS) requirements. Compared
with the cloud servers, edge servers are closer to
the client, which can effectively improve the efficiency
of DNN offloading and reduce the pressure
on network bandwidth. However, edge servers still
suffer from limited computing capacity and insufficient
storage.
Generally, it takes a long time to upload and
execute the DNNmodel, due to the large size of the
DNN model, such as YOLO and FaceNet. Accordingly,
DNN query, i.e., the process of DNN inference,
will only be implemented on the client until
the DNN model is uploaded. However, the client's
computing capacity is low, which leads to the poor
performance of the DNN query. Thus, we perform
the DNN query while uploading the DNN partition,
and then use edge/cloud servers to optimize query
performance. We usually use DNN migration to
achieve two goals: one is to determine the DNN
layer distribution on the edge/cloud server; the
other is to determine the upload order of the DNN
layer. Thus, how to obtain the optimal DNN migration
scheme has become an urgent problem to be
solved. In general, the migration of DNN models
faces the following challenges.
1) Large-scale DNN partition: In order to determine
the upload order of the DNN layer, previous
studies, such as He et al.'s work,2 have to
divide the DNN model, but the DNN partition
is usually very large, so there is no chance to
realize more efficient DNN query performance
throughmore fine-grained partitions.
2) Poor search ability: Most studies use traditional
algorithms, e.g., swarm intelligence algorithms,
to divide DNNmodels and obtain the DNN layer
distribution under each server. However, these
algorithms suffer from poor global search capabilities
and tend to fall into local optimality.
3) Simple computing environment: DNN migration
is only suitable for very simple environments,
not formore realistic computing environments,
such as multiuser multiedge/cloud server environments.
In addition, it is also difficult to
achieve parallel processing of DNNmodels.
May/June 2023
29

IEEE Consumer Electronics Magazine - May/June 2023

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