Advances, Systems and Applications
From: A cloud-edge collaborative task scheduling method based on model segmentation
Notations | Description |
---|---|
\(num\) | the amount of data uploaded by users to the edge |
\(N\) | the number of layers in a neural network model |
\(T_{\delta }\) | task calculation delay on the cloud or edge |
\(\delta \in \{ e,c\}\) | edge node and cloud node respectively |
\(T_{total}\) | transmission delay of intermediate data |
\(T_{c}\), \(T_{e}\) | cloud computing delay, edge computing delay |
\(F\) | FLOPs of the neural network model |
\(F_{c}\), \(F_{e}\) | cloud computing and edge computing capability |
\(F_{\delta }\) | FLOPS of the edge or cloud server |
\(F = \{ f_{1} ,f_{2} ,...,f_{n} \}\) | FLOPs of each layer of the neural network model |
\(C_{in}\) | the number of input characteristic matrices |
\(K_{w}\), \(K_{h}\) | width and height of convolution kernel |
\(C_{out}\) | the number of output characteristic matrices |
\(w\), \(h\) | width and height of output characteristic matrices |
\(FLOPs_{{\text{cov}}}\), \(FLOPs_{fc}\) | FLOPs of convolution and full connection layer |
\(N_{In}\), \(N_{Out}\), | input features, output features |
\(N_{core}\), \(H_{{\text{c}}}\) | cores number and frequency of the processor |
\(N_{float}\) | floating-point operations per cycle of the processor |
\(T_{e} = \{ t_{e,1} ,t_{e,2} ,...,t_{e,n} \}\) | computing delay of each layer at the edge node |
\(T_{c} = \{ t_{c,1} ,t_{c,2} ,...,t_{c,n} \}\) | computing delay of each layer at the cloud node |
\(V_{trans}\), \(V_{up}\),\(V_{down}\) | transmission rate, uplink rate, and downlink rate |
\(O\) | the size of data to be transmitted |
\(O = \{ o_{0} ,o_{1} ,...,o_{n - 1} \}\) | the data output of each layer |
\(T_{up} = \{ t_{up,0} ,t_{up,1} ,...,t_{up,n - 1} \}\) | uplink delay of the output data of each layer |
\(T_{down} = \{ t_{down,0} ,t_{down,1} ,...,t_{down,n - 1} \}\) | downlink delay of the output data of each layer |
\(T_{trans} = \{ t_{trans,0} ,t_{trans,1} ,...,t_{trans,n - 1} \}\) | transmission delay of the output data of each layer |
\(D_{size}\) | the space occupied by the corresponding data type |
\((\dim_{1} ,\dim_{2} ,...,\dim_{m} )\) | the Tensor size output |
\(T_{total} = \{ t_{total,0} ,t_{total,1} ,...,t_{total,n - 1} \}\) | the transmission delay of the data of each layer |
\(T = \{ t_{0} ,t_{1} ,...,t_{{\text{N}}} \}\) | task completion time |
\(M_{total}\) | the total memory of the system |
\(M_{{{\text{wait}}}}\) | system memory consumption without task execution |
\(M = \{ m_{0} ,m_{1} ,...,m_{n} \}\) | memory consumption of each layer when running task |
\(M_{cost} = \{ m_{\cos t,0} ,m_{\cos t,1} ,...,m_{\cos t,n} \}\) | memory consumption of each layer |
\({R}_{memory}=\{{r}_{m,0},{r}_{m,1},...,{r}_{m,n}\}\) | memory occupancy of each layer |
\(split\) | the location of the segmentation point |