IEEE Consumer Electronics Magazine - May/June 2023 - 35

partition P3 to the edge/cloud server, and then
sends CR of the next DNN partition P3 back to
the client. Repeat the uploading process until
the last DNN partition P3 is uploaded to the
edge/cloud server.
Query Thread Consider performing DNN queries
when uploading DNN partitions P3. The client will
repeatedly perform DNN queries, and immediately
propose a new query after the previous query is
completed. Before the DNN model is completely
uploaded, the query can be executed jointly
through the DNN partition P2 under the client and
the DNN partition P3 under the edge/cloud server.
The query thread is the process of uploading
DNN models, and performs a DNN query on the
incomplete uploaded DNN model under the edgecloud
collaborative computing environment. The
local device obtains which DNN partitions have
been uploaded to the edge/cloud server by checking
whether the CR of each DNN partition P3 has
arrived. When a DNN query is triggered, the query
thread will execute the DNN partitions P3 in serial
order. The first local DNN partition is performed,
and then the input matrix (i.e., the output matrix
of the first local DNN partition) and the index of
the second DNN partition P3 are sent to the designated
client or edge/cloud server. After the second
DNN partitions P3 is executed, the input
matrix and the index of the next DNN partitions P3
will be sent to the client or the edge/cloud server
where the next DNN partitions P3 is located. In
this way, the client and the edge/cloud server
jointly execute the DNN partition P3 until the
whole DNNmodel is executed.
IMPROVEMENT AND COMPARISON
Improvement of Partition-Uploading Plan
It is known that the existing partition-uploading
algorithm is only suitable for the situation
from a single client to a single edge/cloud server,
which is obviously not suitable for DNN uploads
in real scenarios. In this part, we will discuss the
performance of the DNN partition-uploading
plan in a multiuser edge-cloud collaborative
computing environment. It is mainly divided into
the following two steps.
1) We simply apply the PSO algorithm to
obtain the corresponding offloading position of
May/June 2023
the DNN subtasks in a multi-user edge-cloud collaborative
environment. 2) According to the
existing partition-upload algorithm, combined
with the unloading position of the DNN subtask,
the DNN partition P1 is divided to obtain the
DNN partition P3, and the upload order of each
DNN partition P3 can be determined.
Experimental Comparison
Experimental Setup We build an edge-cloud
collaborative computing environment R ¼fr1;
r2; ... ;r12g, where the first two belong to the clients,
the last five belong to the cloud servers,
and the remaining five belong to the edge servers.
We set the bandwidth between the local and
the edge is 10 MB/s, between the local and the
cloud is 0.5 MB/s, between the edge and the
cloud is 0.5 MB/s, between different cloud servers
is 5 MB/s, and between different edge servers
is 10 MB/s. The CPU processing capacity
of the client, the edge server, and the cloud
server are set to 1:1 2:3 GHz, 4:2 18:3 GHz,
and 40 120 GHz, respectively. Moreover, the
model size of Alexnet is 223 MB, and the model
size of the VGG model is 548 MB.
Performance Comparison and Analysis
We apply the classic Direct Uploading algorithm,
IONN algorithm, and Recursive Efficiency algorithm
to the multiuser edge-cloud collaborative
computing environment, so as to illustrate the
feasibility and necessity of the DNN partitionuploading
plan in the multiserver environment.
As shown in Figure 5, we track the system
delay changes of repeated queries of the Alexnet
model and VGG model, respectively. Among
them, the horizontal axis represents the upload
delay of the DNN partition P3, and the vertical
axis represents the change of the system delay
with the upload of the DNN partition P3. We consider
performing DNN queries while uploading
DNN partitions. In this process, we assume that
the client will repeatedly put forward DNN
queries, and execute new queries immediately
after the previous one is completed, until the
entire model is uploaded. In addition, During the
upload period of the DNN partition P3, the DNN
system delay will not change during the upload
period of the DNN partition P3.
35
IEEE Consumer Electronics Magazine - May/June 2023

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