Offloading application modules to edge nodes in an edge network
The method dynamically optimizes sub-module offloading in edge computing systems by constructing offload patterns and selecting edge nodes based on real-time conditions, addressing latency and energy consumption issues in edge networks, and enhancing resource utilization.
Patent Information
- Application Number
- PCT/IB2023/054941
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-05-12
- Publication Date
- 2024-11-21
AI Technical Summary
Current edge computing systems face challenges in dynamically optimizing the offloading of application sub-modules to edge nodes due to limited computation resources and rapidly changing network conditions, leading to increased service latency, cost, and energy consumption, especially for real-time and mobile applications with varying quality-of-service requirements.
A method is introduced that constructs a feasible sub-module offload pattern matrix and dynamically selects which sub-modules to offload to specific edge nodes based on real-time operational conditions, using a system comprising module partitioning, offload pattern construction, policy enforcement, observability, monitoring, and optimization engines, including machine learning models, to minimize latency and energy consumption while satisfying quality-of-service requirements.
This approach enables real-time and dynamic offload deployment, reducing service latency and cost for users and improving resource utilization in edge networks, achieving more optimal deployment architectures with sub-second overhead.
Smart Images

Figure IB2023054941_21112024_PF_FP_ABST
Abstract
Description
OFFLOADING APPLICATION MODULES TO EDGE NODES IN AN EDGE NETWORK TECHNICAL FIELD
[0001] Disclosed are embodiments related to the offloading of application sub-modules to edges node in an edge network. BACKGROUND
[0002] Real-time and mobile applications demand a diverse set of services from a network and / or service provider, and each application may have a different quality-of-service (QoS) requirement. These differing QoS requirements provide a challenge for mobile network operators (MNOs) (e.g., operators of 5G networks).
[0003] Edge computing is an emerging technique that aims to meet the increasing computation demands arising from the billions of devices connected to the internet. Furthermore, offloading modules of an application to the edge of the network can improve latency of the application since it can process data closer to user equipment (UE).
[0004] Distributing application modules to the edge can be achieved if applications are designed as a composition of multiple modules. One or more application modules can be brought to the edge of the network which may be geographically and network wise close to end-users.
[0005] The number of modules that will move to the edge will be based on the availability of edge resources and network conditions at any given time. Therefore, it will be important to figure out how and from what edge node offloading service should be requested. Another important issue is the cost of edge network service for the UEs. The UEs desire to get the service within their cost limitation. When the edge node is occupied by the users with high QoS requirements, the edge node prioritizes service provision for the UE with high QoS requirements, therefore, the services for the UEs with middle and low range QoS requirements are prolonged which results in more service latency and service cost to the UEs as well as high energy consumption to the edge system.
[0006] An ad-hoc distribution of modules across the cloud and edge can be detrimental to the performance of an application and in turn affects the QoS. Furthermore, currently there is nosolution from vendors to offload the modules directly from the UE to the edge system. As explained in references [1], [2], and [3], efficient computing offloading and resource allocation have become important issues in edge computing. In addition, due to the randomness and dynamics of edge networks, the long execution time of applications results in more service latency and high service cost to the UE, and also high energy consumption and less efficient resource utilization to the edge network, as explained in reference [4] and [5]. SUMMARY
[0007] Certain challenges presently exist. For instance, existing research for offloading modules from the cloud to the edge is based on static distribution or scheduling of the application by relying on an initial deployment. As for the real-time and mobile applications, an offload module may not completely be served by one edge node due to the unique characteristics of the edge network (i.e., limited computation resources and rapid variation of edge network condition), therefore, there should be a mechanism for real-time offload optimization to dynamically select an appropriate deployment plan, and determine what modules (e.g. sub-modules) should be offloaded to what edge node in a distributed edge network while satisfying QoS requirements of the offload module. In short, there presently does not exist a solution to address the important aspects of real-time offload optimization.
[0008] Accordingly, in one aspect there is provided a method for offloading sub-modules to edge nodes of an edge network. The method includes constructing a feasible sub-module offload pattern, FSOP, matrix for a set of offload sub-modules. The method also includes choosing a set of one or more offload sub-modules to be assigned to a specific edge node for offloading. The method also includes updating a sub-module offload pattern, SOP, matrix based on the FSOP matrix and the set of offload sub-modules chosen to be assigned to the edge node for offloading. The method further includes offloading the set of offload sub-modules in accordance with the SOP matrix.
[0009] In another aspect there is provided a computer program comprising instructions which when executed by processing circuitry of a computing apparatus causes the computing apparatus to perform any of the methods disclosed herein. In one embodiment, there is provided a carrier containing the computer program wherein the carrier is one of an electronic signal, an optical signal, a radio signal, and a computer readable storage medium. In another aspect there is provided a computing apparatus that is configured to perform the methods disclosed herein. The computing apparatus may include memory and processing circuitry coupled to the memory.
[0010] An advantage of the embodiments disclosed herein is that they provide a real-time and dynamic offload deployment mechanism, which leads to a performance improvement and a more optimal deployment architecture. BRIEF DESCRIPTION OF THE DRAWINGS
[0011] The accompanying drawings, which are incorporated herein and form part of the specification, illustrate various embodiments.
[0012] FIG.1 illustrates a system according to an embodiment.
[0013] FIG.2 is a flowchart illustrating a process according to an embodiment.
[0014] FIG.3 is a flowchart illustrating a process according to an embodiment.
[0015] FIG.4 is a flowchart illustrating a process according to an embodiment.
[0016] FIG.5 is a block diagram of a computing apparatus according to an embodiment. DETAILED DESCRIPTION
[0017] An edge network, which comprises one or more edge nodes, has limited computing resources and its network conditions are constantly changing. Accordingly, a module partitioning mechanism should be provided so that the offloading of sub-modules can be completely served by the edge network. The partitioning mechanism specifies what pattern of sub-modules can be served by which edge nodes.
[0018] As the applications deployed and system and network conditions change rapidly, a system is proposed that configures the distribution of offload modular components by learning from the operational environment and feedback of each deployment. The system observes the network and compute conditions on a distributed Edge network and applying this informationwhen offloading from the UE to the distributed Edge network. As the information changes rapidly, it can be said that an Edge network digital twin is created and used to optimize offloading with real-time observed data. Therefore, instead of static distribution and scheduling, the offloaded module deployment can be done in a real-time and dynamic way. The merit of the real-time offloading optimization mechanism is to select an appropriate deployment plan dynamically and determine what sub-modules should be offloaded to what Edge node in a distributed Edge network while satisfying QoS requirements of the offload module. The proposed system has a sub-second overhead which results in less service latency and less service cost for the UE as well as less energy consumption and more efficient resource utilization for the Edge network.
[0019] The proposed system offloads sub-modules to edge nodes in a real-time and dynamic manner. To this aim, a pattern is constructed that indicates how sub-modules should be offloaded to the Edge nodes. It is possible to offload all the sub-modules to one Edge node only. The decision of what sub-module needs to be offloaded to what Edge node is made based on operational conditions such as the availability of resources, Edge network conditions, and QoS requirements as the combination of the computation and communication latencies of the distributed applications. The decision of which components need to be distributed is made based on operational conditions such as the availability of resources, edge network conditions, and QoS requirements as the combination of the computation and communication latencies of the distributed applications.
[0020] FIG.1 illustrates a system 100 according to an embodiment for offloading application modules to one or more edge nodes of an edge network. In one embodiment, all edge nodes are in a base stations or co-located with a component of a base station.
[0021] Because an edge node may be restricted due to limited resources and the availability of resources changes over time due to varying network and system conditions, system 100 determines how and what edge node should be selected for elastic edge applications (i.e., to serve offload module).
[0022] To this aim, in one embodiment the system includes the following components: (1) a module partitioning function (MPF) 102; (2) an offload pattern constructor (OPC) 104; (3) a policy engine (PE) 106; (4) an observability function (OF) 108; (5) a monitor 110; (6) a sub-module offloader (SMO) 112; and (7) an optimization engine (OE) 114 (e.g., a trained machine learning (ML) model or a trained artificial intelligence (AI) model).
[0023] MPF 102 partitions a module into sub-modules and indicates whether a particular sub-module should be considered as an offload sub-module or not. The MPF denotes “1” for the sub-modules to be offloaded and “0” otherwise. Then, MPF 102 constructs a sub-module offload pattern where the rows of the pattern correspond to the sub-modules indexes and the columns correspond to the feasible sub-module offload pattern (FSOP). An example for four sub-modules is shown below in equation (2).
[0024] OPC 104 constructs an offload pattern which indicates that the sub-modules should be offloaded to what edge node. The rows of this pattern correspond to the indexes of offloading sub-modules and the columns correspond to the feasible edge node. An example is shown in equation (3) below.
[0025] PE 106 allows an edge network provider to create, monitor, and enforce rules about how edge network resources can be accessed. The policy engine relies on a combination of network analytics fed by monitor 110 and programmed rules to grant dynamic permissions based on any number of factors, such as hierarchy of need.
[0026] OF 108 observes the deployment environment in an edge network and collects real-time data.
[0027] Monitor 110 gets data collected by OF 108 and performs analysis using the data. The result of the analysis is set to PE 106 so that PE 106 may update policies and create new policies if needed.
[0028] SMO 112, based on operational parameters derived from monitor 110, such as edge-server workload, available resources, data size of the offload sub-module, and QoS requirements of the UE, decides what sub-modules should be offloaded to what edge node so that the service latency and cost, and energy consumption for computational procedure of offloaded sub-modules are minimized while satisfying the QoS requirements of UE.
[0029] OE 114 gets data from a Real-Time Offload Optimization system 116, analyzes it using AI / ML and might propose new policies for module partitioning and placement of sub- module offloading. It also provides the evaluation by using a digital twin of the edge network fornew proposed models which means what module partitioning model, what sub-module offloading pattern construction, and what placement of sub-module offloading will result in more minimization of service latency and cost for the UE, and energy consumption of edge network if new deployment is executed.
[0030] The following describes an algorithm for a dynamic partitioning and placement for real-time offload optimization mechanism.
[0031] It is assumed that the Edge network 190 consists of ^^^^ = {1, … , ^^^^, … , ^^^^} Edge nodes and ^^^^ = {1, … , ^^^^, … , ^^^^} UEs. Based on operational parameters derived from Edge network such as Edge-server workload, available resources, data size of the offload sub-module, and QoS requirements of the UE, the real-time offload optimization (RLO2) mechanism should obtain a dynamic partitioning and placement for offload decision making.
[0032] It is further assumed that each UE invokes ^^^^^^^^= {1, … , ^^^^^^^^, … , ^^^^^^^^} sub-modules over time. Furthermore, the chronological order of generated sub-modules over the time is kept.
[0033] It is further assumed that each offloading module of the UE has a size of ^^^^^^^^in which ^^^^ indicates the size of module and index ^^^^ indicates the UE to which the module belongs. Suppose that the module of UE ^^^^ is partitioned to four sub-modules. Thus, ^^^^^^^^ ^^^^ ^^^^^^^ ,in which we associate each sub-module of a module of UE ^^^^ with a binary decision variable^^^^ ^^^^, ^^^^ ^^^^{0,1}, ^^^^ ={1, … , ^^^^}which indicates whether a particular sub-module should be offloadednode ^^^^ or not.
[0034] The feasible sub-module offload pattern (FSOP) matrix can be constructed by denoting a sub-module to be offloaded to an Edge node by a “1” and not offloaded by a “0”, thus forming the FSOP matrix ^^^^^^^^, ^^^^for UE ^^^^. The rows of this matrix correspond to the sub-modules indexes and the columns correspond to the feasible sub-module offloading patterns. Therefore, in general, the FSOP matrix including all feasible sub-module offloading patterns for the case ^^^^^^^^= 4 offloading sub-modules can be written as follows:0 1 0 0 0 1 0 0 1 0 1^^^^^^^^=�0 0 1 0 0 1 1 0 1 1 1 2),which is
[0035] It is further assumed that when multiple sub-modules should be offloaded to an Edge node (i.e., the columns 6-11 of Wl,j), these sub-modules should be adjacent to each other. Thus, the offload matrix is of order ^^^^^^^^× ^^^^, in which Lndenotes the number of offload sub- modules, and ^^^^ denotes the total number of feasible sub-module offload patterns given by ^^^^ = 1 2 ×� ^^^^^^^^2+ ^^^^^^^^� + 1, which is 11 in this example. vector of length ^^^^ ^^^^ is denoted as ^^[ ] ^^^^^^ = ^^^^1, … , ^^^^^^^^, where ^^^^^^^^= � ^^^^ , … , ^^^^^^^^^^^,1^^^^^, ^^^^�. Suppose that one has ^^^^^^^^= 4 sub-modules from UE ^^^^ and ^^^^ = 2 Edge nodes. that ^^^^^^^^= 1 is allocated to Edge node ^^^^ = 1 and ^^^^^^^^= 2, 3, ^^^^ ^^^^ ^^^^ 4 are allocatedto Edge node ^^^^ = 2. Thus, the sub-module offloading pattern (SOP) matrix is given as follows:^^^^ ^^^^, ^^^^ = ^^^^ ^^^^ ^^^^ ^^^^, ^^^^ =�0 1 0 0 0 0 0 0 0 0 00 0 0 0 0 0 0 0 0 1 0� ( ^^^^ ^^^^.3),for Edge node ^^^^ = 2 the sub-module offloading pattern ^^^^ = 10 is offloaded which includes the sub- modules ^^^^^^^^= 2, 3, ^^^^ ^^^^ ^^^^ 4 of UE ^^^^.
[0037] Hence, the aim is to construct SOP matrix so that sub-modules are offloaded to Edge in a way that the total latency of UEs and total energy consumption of Edge network including Edge UEs and Edge nodes are minimized while satisfying the QoS requirements of the UEs.
[0038] We denote the queue order of ^^^^^^^^sub-modules generated by ^^^^ Edge UEs in each Edge node ^^^^ as ^^^^^^^^= { ^^^^1, … , ^^^^^^^^}^^^^, in which ^^^^^^^^will be a matrix of length ^^^^ ^^^^ and ^^^^^^^^= the energy consumption and latency to complete the ^^^^th sub-module^^^^^^^^^^^^= ^^^^^^^^^^^^, ^^^^( ^^^^ ^^^^ ^^^^{ ^^^^^^^^^^^^^^^^ ^^^^, ^^^^ ^^^^, ^^^^^^^^^^^^ ^^^^ ^^^^, ^^^^}+ ^^^^^^^^^^^^ ^^^^ ^^^^, ^^^^)+�1 − ^^^^^^^^^^^^, ^^^^� ^^^^^^^^^^^^ ^^^^ ^^^^, ^^^^(^^^^ ^^^^.4),^^^^^^^^^^^^ = ^^^^ ^^^^^^^^, ^^^^� ^^^^ ^^^^^^^^ ^^^^ ^^^^, ^^^^ ^^^^ ^^^^^^^^ ^^^^ ^^^^, ^^^^ + ( ^^^^ ^^^^ ^^^^� ^^^^ ^^^^^^^^ ^^^^ ^^^^, ^^^^ ^^^^ , ^^^^ ^^^^^^^^ ^^^^ ^^^^, ^^^^� − ^^^^ ^^^^^^^^ ^^^^ ^^^^, ^^^^� ^^^^ ^^^^^^^^ ^^^^ ^^^^, ^^^^ + ^^^^ ^^^^^^^^ ^^^^ ^^^^, ^^^^ ^^^^ ^^^^^^^^ ^^^^ ^^^^, ^^^^ )−^^^^^^^^^^^^^^ ,of^^^^^UE ^^^^, ^^^^^^^^=^^^ ^^^^ ^^^^ ^^^^is the^^^^^^^^ ^^^^ ^^^^, ^^^^ ^^^^ ^^^^ ^^^^ ^^^^server computing time, ^^^^ ^^^^ ^^^^ ^^^^ ^^^^, ^^^^is the time used to transmit sub- module(s) to
[0039] The size of the sub-modules offloaded to Edge node ^^^^ is: ^^^^^^^^^^^^=� ^^^^^^^^^^^^, ^^^^^^^^^^^^^^^^( ^^^^ ^^^^.6),node ^^^^. Therefore, the computational time of sub-modules with size ^^^^^^^^in Edge node ^^^^ is as follows: ^^^^ ^^^^^^^^^^^^ ^^^^, ^^^^ ^^^^ =^^^^ ^^^^ ^^^^ ^^^^ ^^^^^^^(^^^^ ^^^^.7).^^^^^ ^^^^ ^^^^ ^^^^
[0040] Hence, the computational time of a sub-module with size ^^^^^^^^received from UE ^^^^ is: ^^^^^^ ^^^^ ^^^^^^^^^^^^^^^ ^ ^^^^ ^^^^ ^^^^^^ ^^^^, ^^^^ =^^^(^^^^ ^^^^.8).^^^^^ ^^^^ ^^^^ ^^^^
[0041] Furthermore, the computational energy for a sub-module with size ^^^^^^^^is: ^^^^^^^^ ^^^^ ^^^^, ^^^^ ^^^^= ^^^^ ^^^^^^^^^^^^ ^^^^^^^2^^^^^ ^^^^ ^^^^(^^^^ ^^^^.9),computational capacity of Edge server. Also, the computational energy and power for sub-modules with size ^^^^ ^^^^ which is under processing in Edge server ^^^^ are as follows:^^^^ ^^^^ ^^^^ ^^^^, ^^^^ ^^^^ = ^^^^ ^^^^ 2^^^^ ^^^^ ^^^^ ^^^^ ^^^ ^^^^^ ^^^^ ^^^^ ^^^^ ( ^^^^ ^^^^.10),).
[0042] All UEs are randomly distributed in the plane and move in random directions with speed ^^^^^^^^= { ^^^^1, … , ^^^^^^^^}. The upload transmission delay for each bit of a sub-module ^^^^th with size ^^^^^^^^of UE ^^^^th is: ^^^^^^^ ^^^^ ^^^^^^^^^ ^^^^ ^^^^, ^^^^=^^^^ log ^^^^, ( ^^^^ ^^^^.12),2�1 + ^^^^ ^^^^ ^^^^ ^^^^ ℎ⁄ ^^^^2�channel gain ℎ follows the free space path loss model. Generally, the size of the computing result of a sub-module is much smaller than the sub-module size thus, the latency of Edge server transmitting the computing results to a UE can be ignored.
[0043] Sub-module Scheduling Model
[0044] The length of the server’s waiting queue is assumed to be ^^^^^^^^, so one must sort up to ^^^^ ^^^^ sub-modules. If the waiting queue is not full and there are ^^^^ sub-modules in the queue(^^^^ < ^^^^ ^^^^) then there are ^^^^ + 3 options for the server to handle the sub-module offload request:1) reject it; 2) place it in front of the existing ^^^^ sub-modules, or 3) place it in the last position of the queue. If the newly arrived sub-module is placed in the front of the queue, it will increase the processing time of the existing ^^^^ sub-modules. Assume that the processing latency required by the original ^^^^ sub-modules is ^^^^^^^^,1, … , ^^^^^^^^, ^^^^the processing latency of the original sub-modules becomes ^^^^^^^^,1, … , ^^^^^^^^, ^^^^after a new sub-module arrives, then we define the loss ^^^^^^^^caused by the new sub-module to the system as ^^^^^^^^=∑ ^^^^^^^=^ ^^^^ ^^^^, ^^^^− ^^^^ ^^^^, ^^^^1^^^^ ^^^^.
[0045] The goals of the sub-algorithms are to minimize 1) weighted sub-module processing time and 2) weighted energy consumption to complete the sub-module computing process. Because the optimization goals are coupled with { ^^^^, ^^^^}, it is impossible to optimize them independently. A trade-off is made between these two goals and define the problem as follows: ^^^^1: m ^^^^, ^i^^n ^ ^^^^( ^^^^, ^^^^) = ^^^^ ^^^^( ^^^^, ^^^^) + ^^^^ ^^^^( ^^^^, ^^^^) ( ^^^^ ^^^^.13 ^^^^), ^^^^. ^^^^ ^^^^ ^^^^, ^^^^ ^^^^ {0, 1} ( ^^^^ ^^^^.13 ^^^^),^^^^ ^^^^ ^^^^ {1, 2, … , ^^^^} ( ^^^^ ^^^^.13 ^^^^),where ^^^^ and ^^^^ are weighting parameters, ^^^^(^^^^, ^^^^)=∑^^^^^^^^^^^^ ^^^^− ^^^^ ^^^^and ^^^^(^^^^, ^^^^)=∑ ^^^^ ^^^^^^^^ ^^^^ .
[0046] The objective function is to minimize the weighted sum of the system execution time and the energy consumption. The constraint (13 ^^^^) guarantees a sub-module can only be processed locally or in the Edge server. The constraint (13 ^^^^) guarantees the number of waiting sub-modules will not exceed ^^^^. However, this problem cannot be solved efficiently since ^^^^1 is NP-hard problem. Thus, a solution is proposed to use AI deep reinforcement learning to solve it.
[0047] Algorithm 1: Sub-module Offload Pattern Construction Matrix for UEs in Edge Network.
[0048] The table below illustrates pseudo code for generating a SOP matrix. 1: Initialize ^^^^ ^^^^{1, 2, … , ^^^^}, ^^^^ ^^^^{1, 2, … , ^^^^}, ℒ ^^^^ ={1, … , ^^^^ ^^^^}∀ ^^^^ ^^^^ ^^^^, ^^^^ ^^^^ = ∅ ∀ ^^^^ ^^^^ ^^^^,^^^^^^^^^^^^= ℒ^^^^∀ ^^^^ ^^^^ ^^^^ 2: ∀ ^^^^ ^^^^ ^^^^ 3: ^^^^ ^^^^ ^^^^ ^^^^ = 1: ^^^^ 4: repeat 5: ∀ ^^^^^^^^^^^^ ^^^^^^^^^^^^∩ ℒ^^^^6: ∀ ^^^^ ^^^^ ^^^^ 7: ^^^^ ^^^^ ^^^^ ^^^^^^^^= 1: ^^^^^^^^8: ^^^^ ^^^^ ^^^^ ^^^^ = 1: ^^^^ 9: ^^^^ ^^^^ ^^^^ ^^^^ ( ^^^^^∗^^^, ^^^^∗) = argmin ^^^^ ^^^^^^^^^^^^ ^^^^^ + ^^^^ ^^^^^^^^^^^^ ^^^^ ^^^^ , ^^ ^^^^^^^ ^^^^ ^^ ^^^^ ^^^^, ^^^^ ^^^^ ^^^^10: ^^^^^^^^∗= ^^^^^^^^∪ ^^^^^∗^^^11: ^^^^^^^^^^^^∗= { ^^^^ ^^^^ ^^^^( ^^^^^^^^∗) − 1, ^^^^ ^^^^ ^^^^( ^^^^^^^^∗) + 1} 12: ^^^^^^^^= ^^^^^^^^\ ^^^^^∗^^^13: ^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^ 14: ^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^15: until ^^^^^^^^= ∅ 16: ^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^ 17: Return sub-module offload pattern matrix
[0049] Details of Algorithm 1
[0050] Line 1 initializes ^^^^ UEs having ℒ^^^^sub-modules to offload to the ^^^^ Edge nodes. We note that the feasible sub-modules at the beginning is equal to the total number of sub- modules.
[0051] Line 2 indicates that the steps provided in Lines 3-17 are valid for every UE.
[0052] Line 3 indicates that for ^^^^ = 1: ^^^^ UEs repeat the steps provided in Lines 5-17.
[0053] Line 4 repeats the steps in Lines 5-17.
[0054] Line 5 indicates that for each sub-module in the feasible sub-module the steps provided in Lines 7-17 are valid.
[0055] Line 6 indicates that for each Edge node belongs the total Edge node set ^^^^ the following steps provided in Lines 9-17 are valid.
[0056] Lines 7-9 represent that among ^^^^^^^^= 1: ^^^^^^^^sub-modules of UE ^^^^th and among ^^^^ = 1: ^^^^ Edge nodes the algorithm finds the sub-module ^^^^^∗^^^and Edge node ^^^^∗in which if the sub-module ^^^^^∗^^^is offloaded to Edge node ^^^^∗the total latency of UE ^^^^th ^^^^^^^^ ^^^^and total loss(^^^^^^^^)of the Edge ^^^^th will be minimized.
[0057] Line 10 represents that in the sub-module offloading pattern matrix indicate with “1” that the sub-module ^^^^^∗^^^is offloaded to Edge node ^^^^.
[0058] Line 11that next sub-module for offloading is selected from the sub- module(s) that are adjacent to the previously sub-module indicated by “1” in sub-module offloading pattern.
[0059] Line 12 represents that the sub-module indicated by “1” should be removed from the list of sub-modules to be offloaded.
[0060] Line 13 ends the “for clause” provided in Line 8.
[0061] Line 14 ends the “for clause” provided in Line 7.
[0062] Line 15 indicates that are all sub-modules of UE ^^^^ indicated by “1” offloaded?
[0063] Line 16 indicates that the process is finished.
[0064] Line 17 returns the sub-module offload pattern matrix.
[0065] Algorithm 2: Edge Node Selection for Sub-module Offload Pattern in Distributed Edge Network.
[0066] The pseudo code below illustrates a process for edge node selection. 1: For UE ^^^^ ^^^^ {1, 2, … , ^^^^} covered by ^^^^ ^^^^ {1, 2, … , ^^^^} Edge nodes which has ^^^^ = 1, … , ^^^^ sub-modules for offloading 2: Initialize Loss = 0, order = 0, ^^^^^^^^= 0, ^^^^ ^^^^ = ^^^^ 3: While ^^^^ ^^^^ ≠ ^^^^ do 4: Loss = ^^^^ ^^^^^^^^− ^^^^^^^^ ^^^^ ^^^^+ ^^^^^^^^ ^^^^ ^^^^^^^^5: for ^^^^ = 1, … , ^^^^ do 6: for ^^^^ = 1, … , ^^^^ do 7: if ^^^^ == 1 8: for ^^^^ = 1, … , ^^^^ do 9: obtain ^^^^^^^^10: if Loss > ^^^^ ^^^^^^^^− ^^^^ ^^^^ ^^^^ ^^^^^^^^ ^^^^ ^^^^+ ^^^^�^^^^+ ^^^^^^^^� then 11: Loss = ^^^^ ^^^^^^^^ ^^^^ ^^^^+ ^^^^�^^^^− ^^^^ ^^^^ ^^^^ ^^^^^^^^+ ^^^^^^^^� 12: order = order + 1 13: end if 14: end for15: j = j + 1 16: else 17: rew = − ^^^^(^^^^, ^^^^)− ^^^^ ^^^^ ^^^^18: Q(^^^^, ^^^^, ^^^^)= rew19: ^^^^ = ^^^^ + 1 20: for ^^^^ = ^^^^ to | ^^^^ − ^^^^ + 1| do 21: Update the state space ^^^^ = { ^^^^^^^^, ^^^^^^^^, ^^^^^^^^, ^^^^^^^^, ^^^^} where ^^^^^^^^, ^^^^^^^^, ^^^^^^^^, ^^^^^^^^are the vectors of sub-module’s deadline “d”, the movement direction, the location of the UE, and the size of sub-module respectively. Furthermore, ^^^^ indicates the state information of the sub-module which is currently under processing. 22: take an action randomly from action set; 23: a = ^^^^ 24: calculate the reward ^^^^ = ^^^^ ^^^^ ^^^^ + ^^^^max ^^^^( ^^^^, ^^^^, ^^^^) 25: perform a gradient descent step on �^^^^ − ^^^^(^^^^, ^^^^, ^^^^)�226: end for 27: end if 28: end for 29: end for
[0067] Details of Algorithm 2
[0068] Line 1 indicates that for each UE ^^^^ ^^^^{1, 2, … , ^^^^}covered by Edge node ^^^^ ^^^^{1, 2, … , ^^^^}which has ^^^^ = 1, … , ^^^^ sub-modules for offloading follow the steps provided in Lines 2-29.
[0069] Line 2 initializes the Loss = 0, order = 0, ^^^^^^^^= 0, and the total sub-modules offload pattern ^^^^ ^^^^ = ^^^^.
[0070] Note that “Loss” calculates the term (13) which is related to local and Edge node processing. However, ^^^^^^^^is calculated when a new sub-module arrives and the processing latency that it causes to the Edge node system.
[0071] Line 3 indicates that while the algorithm has not reached to the total sub-module offload pattern ^^^^, repeat the steps provided in Lines 4-29.
[0072] Line 4 calculates the Loss in local processing on device.
[0073] Line 5 indicates that for all Edge node ^^^^ = 1, … , ^^^^ do the steps provided in Lines 6-29.
[0074] Line 6 indicates that for all sub-module offload patterns ^^^^ = 1, … , ^^^^ do the steps provided in Lines 7-29.
[0075] Line 7 represents that if the sub-module offload pattern has been indicated with “1” in the sub-module offload pattern matrix then, do the steps provided in Lines 8-15. Otherwise, go to step 17.
[0076] Line 8 indicates that for all original sub-modules ^^^^ = 1, … , ^^^^ (i.e., original sub- modules refer to the sub-modules before the new sub-module arrives to the Edge system) do the step provided in Line 9.
[0077] Line 9 calculates the value of ^^^^^^^^.
[0078] Line 10 indicates that if Loss due to local processing on device is greater than the Edge system then, the sub-module should be offloaded to the Edge system. Furthermore, the value of Loss should be replaced with the loss due to the sub-module offload processing in the Edge node.
[0079] Line 11 calculates the Loss in Edge system due to the processing of sub-module offload pattern.
[0080] Line 12 indicates that the order number should be increased by “1” due to the arrival of new sub-module.
[0081] Line 13 ends the “if clause” provided in Line 10.
[0082] Line 14 ends the “for clause” provided in Line 8.
[0083] Line 15 indicates that go to the next sub-module offload pattern in SOP matrix. This loop is repeated until the algorithm reaches to ^^^^ = ^^^^ which ^^^^ indicates the total sub-module offload pattern in SOP matrix.
[0084] Line 16 indicates that if the statement in Line 7 is not valid then it goes to Line 17.
[0085] Line 17 calculates the reward.
[0086] Line 18 replaces the value of reward (i.e. the “rew” in Line 17) with q-function in deep reinforcement learning (DRL).
[0087] Line 19 represents that do the steps provided in Lines 6-18 for all ^^^^ Edge nodes.
[0088] Line 20 represents the action space in DRL method in which the total number of actions is equal to ^^^^, the total number of Edge nodes for sub-module offloading.
[0089] Line 21 represents the state space in DRL method that includes ^^^^ ={ ^^^^ ^^^^, ^^^^ ^^^^, ^^^^ ^^^^ , ^^^^ ^^^^, ^^^^} where ^^^^ ^^^^, ^^^^ ^^^^, ^^^^ ^^^^ , ^^^^ ^^^^ are the vectors of sub-module’s deadline “d”, the movementdirection, the location of the UE, and the size of sub-module respectively. Furthermore, ^^^^ indicates the state information of the sub-module which is currently under processing.
[0090] Line 22 represents that take an action randomly from action set.
[0091] Line 23 represents that replace the value of “ ^^^^” in “ ^^^^” which indicates the action in DRL method.
[0092] Line 24 calculates the reward ^^^^ = ^^^^ ^^^^ ^^^^ + ^^^^max ^^^^( ^^^^, ^^^^, ^^^^).
[0093] Line 25 performs a gradient descent step on� ^^^^ − ^^^^(^^^^, ^^^^, ^^^^)�2to minimize it.
[0094] Line 26 ends the “for clause” provided in Line
[0095] Line 27 ends the “if clause” provided in Line 7.
[0096] Line 28 ends the “for clause” provided in Line 6.
[0097] Line 29 ends the “for clause” provided in Line 5.
[0098] FIG.2 is a flow chart illustrating a process 200, according to an embodiment, for offloading sub-modules of a UE 101. The process illustrated by the flowchart can be used for any number of UEs. Process 200 may begin in step s202.
[0099] Step s202 comprises constructing a feasible sub-module offload pattern, FSOP, matrix for a set of offload sub-modules. Step s204 comprises choosing a set of one or more offload sub-modules to be assigned to a specific edge node for offloading. That is, for example, the set of offload sub-modules choses in step s204 satisfies the decision metric ^^^^ ^^^^^^^^^^^^ ^^^^^^^^+ ^^^^ ^^^^^^^^.
[0100] Step s206 comprises updating a sub-module offload pattern, SOP,on the FSOP matrix and the set of offload sub-modules chosen to be assigned to the edge node for offloading. Step s208 comprises offloading the set of offload sub-modules in accordance with the SOP matrix.
[0101] FIG.3 is a flow chart illustrating a process 300, which is one specific implementation of process 200. Process 300 can be used for any number of UEs. Process 300 may begin in step s301.
[0102] Step s301 comprises creating an FSOP matrix (see equation (2) for an example FSOP matrix). As described above, the FSOP matrix is a matrix of order ^^^^^^^^× ^^^^ and each column of the FSOP matrix identifies a set of one or more adjacent sub-modules (i.e., it identifies either a single sub-module or a set of two or more adjacent sub-modules). Using the FSOP matrix shown in equation (2) as an example, column 8 of the FSOP matrix identifies a set of sub-modules consisting of sub-modules {3,4}, and column 9 of the FSOP matrix identifies a set of sub- modules consisting of sub-modules {1, 2, 3}.
[0103] Step s302 comprises an initialization step (this step corresponds to line 1 of algorithm 1). To illustrate the process it is assumed that UE 101 invokes four sub-modules (i.e., Ln= {1, 2, 3, 4}) and edge network 190 includes two edge nodes (i.e., m={1,2}). Variables ln, ln*, and temp[] are initialized also. As shown in step 202, lnis initially set equal to Ln, ln* is initially set equal to {} (i.e., the NULL set), and temp[] is also initialized to the NULL set. The temp array allows keeping track of which sub-modules have been assigned to which edge node.Each element of temp[] is a vector (i.e., temp[] is a matrix), and in this case because there are two edge nodes, temp[] consists of two vectors: temp[1] which will contain the list of sub- modules assigned to edge node 1 and temp[2] which will contain the list of sub-modules assigned to edge node 2.
[0104] Step s304 comprises determining, from the set of sub-modules identified by ln, an optimal set of sub-module(s) to offload and the edge node to which the sub-module(s) are assigned for offloading. After step s304 is performed, the set ln* contains the indices of the determined optimal sub-modules and m* is equal to the index of the selected edge node. As an example, ln* = {2},and m* =1, which means that sub-module 2 has been chosen to be assigned to edge node 1 for offloading. The first time that step s304 is performed, it is noted that lnis equal to Ln(i.e., all of the UE’s sub-modules).
[0105] Step s306 comprises updating the temp[] array. Specifically, temp[m*] = temp[m*] Ս ln*. The first time in the process that step s306 is executed temp[m*] will simply equal ln* because temp[m*] was initialized to NULL.
[0106] Step s308 comprises removing the elements of ln* from ^^^^^^^^(i.e., in set notation, Ln= Ln / ln*).
[0107] Step s310 comprises determining the set of sub-modules that are adjacent to the sub-modules included in ln*. More formally, this set of adjacent sub-modules is denoted Lfand Lf= {min(ln*) -1, max(ln*) + 1}). Hence, if ln* = {2}, then Lf= {1,3}.
[0108] Step s312 comprises setting lnequal to the intersection of Lfand Ln(i.e., Ln= LfՈ Ln), which has the effect of setting lnequal to Lfand then removing from lnany sub-module index that is not in set Ln. For instance, if Lf= {2,4} and Ln= {1,3}, then lnwill be an empty set. In contrast, if Lf= {1,4} due to ln* = {2,3} and Ln= {1,3}, then lnwill be {1}.
[0109] In step s314 it is determined whether lnis an empty set. If it is not, the process goes back to step s304, otherwise the process proceeds to step s316.
[0110] Step s316 comprises generating a SOP matrix based on the FSOP matrix and temp[]. As described above, the SOP matrix is of order ^^^^ × ^^^^, where M is the number of available edge nodes in the edge network. That is, each row of the SOP matrix corresponds to one of the available edge nodes and each element in the row identifies a column from the FSOPmatrix (as noted above, each column of the FSOP matrix identifies a set of one or more adjacent sub-modules). As further described above, element 1 of temp[] will contain the list of sub- modules, if any, assigned to be offloaded to edge node 1 and element 2 of temp[] will contain the list of sub-modules, if any, assigned to be offloaded to edge node 2.
[0111] Accordingly, generating the SOP matrix based on the FSOP matrix and temp[] comprises: (1) initializing the SOP matrix by setting each element of the matrix to 0 and (2) for each edge node, determining the column of the FSOP matrix that identifies the sub-modules assigned to the edge node (i.e., the sub-modules in temp[m], where m is the index of the edge node) and placing a “1” in the SOP matrix at position m,c where c is the column of the FSOP matrix that identifies the sub-modules assigned to the edge node. As a specific example, assuming the FSOP matrix shown in equation (2), if temp[1]={3,4} and temp[2]={1,2}, then a 1 is placed at positions SOP1,8and SOP1,6because column 8 of the FSOP matrix identifies sub- modules {3,4}. That is, using the example, the SOP matrix will appear as shown below: ^^^^ ^^^^ ^^^^ =�0 0 0 0 0 0 0 1 0 0 000 0 0 0 1 0 0 0 0 0� .
[0112] Step s318 comprises offloading sub-modules in accordance with the SOP. For example, using the SOP shown above, because in the first row of the SOP there is a 1 in position 8, sub-modules 3 and 4 are offloaded to edge node 1, and, because in the second row of the SOP there is a 1 in position 6, sub-modules 1 and 2 are offloaded to edge node 2. Hence, in one embodiment, edge node 1 invokes sub-modules 3 and 4 and then provides the results of the execution of those sub-modules to UE 101 and edge node 2 invokes sub-modules 1 and 2 and then provides the results of the execution of those sub-modules to UE 101. In this way, sub- modules are offloaded.
[0113] FIG.4 illustrates an alternative process in which the SOP is updated on each iteration (i.e., the temp[] array is not needed). More specifically, step s306 is replaced with step s406. Step s406 comprises updating the SOP based on the FSOP, the current state of the SOP, ln*, and m*. For example, based on the current state of the SOP and ln* one can determine the set of sub-modules currently assigned to edge node m*. Then, each position m*,i (i=1 to J) of the SOP is set to 0, and, based on the set of sub-modules currently assigned to edge node m*, the column of the FSOP matrix that identifies the sub-modules assigned to edge node m* isdetermined and a “1” is placed in the SOP matrix at position m*,c where c is the column of the FSOP matrix that identifies the sub-modules assigned to the edge node m*. As a specific example, if m* = 1, ln* = {2,3} and the current state of the SOP is as shown below: ^^^^ ^^^^ ^^^^ =�0 1 0 0 0 0 0 0 0 0 000 0 0 1 0 0 0 0 0 0� , then the current state of the SOP indicates that sub-module 1 has been assigned to edge node 1 and, hence, the sub-modules currently assigned to edge node 1 is: {1} Ս ln*, which is equal to {1,2,3} in this particular example because ln* = {2,3}. Assuming the FSOP matrix shown in equation (2), the column of the FSOP matrix that identifies sub-modules {1, 2, 3} is column 9. Hence, the SOP is updated to a new state such that position m*,9 is set to 1 and all otherpositions in row m* are set to 0, as shown below:^^^^ ^^^^, ^^^^ =�0 0 0 0 0 0 0 0 1 0 00 0 0 0 1 0 0 0 0 0 0� .
[0114] In this manner, the SOP is updated after each iteration.
[0115] FIG.5 is a block diagram of computing apparatus 500, according to some embodiments. Computing apparatus may be a component of UE 101. As shown in FIG.5, computing apparatus 500 may comprise: processing circuitry (PC) 502, which may include one or more processors (P) 555 (e.g., one or more general purpose microprocessors and / or one or more other processors, such as an application specific integrated circuit (ASIC), field- programmable gate arrays (FPGAs), and the like), which processors may be co-located in a single housing or in a single data center or may be geographically distributed (i.e., computing apparatus 500 may be a distributed computing apparatus); at least one network interface 548 (e.g., a physical interface or air interface) comprising a transmitter (Tx) 545 and a receiver (Rx) 547 for enabling computing apparatus 500 to transmit data to and receive data from other nodes connected to a network 110 (e.g., an Internet Protocol (IP) network) to which network interface 548 is connected (physically or wirelessly) (e.g., network interface 548 may be coupled to an antenna arrangement comprising one or more antennas for enabling computing apparatus 500 to wirelessly transmit / receive data); and a storage unit (a.k.a., “data storage system”) 508, which may include one or more non-volatile storage devices and / or one or more volatile storage devices. In embodiments where PC 502 includes a programmable processor, a computer readable storage medium (CRSM) 542 may be provided. CRSM 542 may store a computer program (CP)543 comprising computer readable instructions (CRI) 544. CRSM 542 may be a non-transitory computer readable medium, such as, magnetic media (e.g., a hard disk), optical media, memory devices (e.g., random access memory, flash memory), and the like. In some embodiments, the CRI 544 of computer program 543 is configured such that when executed by PC 502, the CRI causes computing apparatus 500 to perform steps described herein (e.g., steps described herein with reference to the flow charts). In other embodiments, computing apparatus 500 may be configured to perform steps described herein without the need for code. That is, for example, PC 502 may consist merely of one or more ASICs. Hence, the features of the embodiments described herein may be implemented in hardware and / or software.
[0116] While various embodiments are described herein, it should be understood that they have been presented by way of example only, and not limitation. Thus, the breadth and scope of this disclosure should not be limited by any of the above-described exemplary embodiments. Moreover, any combination of the above-described elements in all possible variations thereof is encompassed by the disclosure unless otherwise indicated herein or otherwise clearly contradicted by context.
[0117] Additionally, while the processes described above and illustrated in the drawings are shown as a sequence of steps, this was done solely for the sake of illustration. Accordingly, it is contemplated that the order of the steps may be re-arranged, and some steps may be performed in parallel. Also, as used herein “a” means “at least one” or “one or more.”
[0118] References
[0119] [1] T. Mengistu, A. Alahmadi, A. Albuali, Y. Alsenani, and D. Che, ‘‘A ‘no data center’ solution to cloud computing,’’ in Proc. IEEE 10th Int. Conf. Cloud Comput. (CLOUD), Jun.2017, pp.714-717.
[0120] [2] T.-Y. Kan, Y. Chiang, and H.-Y. Wei, ‘‘Task offloading and resource allocation in mobile-edge computing system,’’ in Proc.27th Wireless Opt. Commun. Conf. (WOCC), Apr.2018, pp.1–4.
[0121] [3] X. Wei, S. Wang, A. Zhou, J. Xu, S. Su, S. Kumar, and F. Yang, ‘‘MVR: An architecture for computation offloading in mobile edge computing,’’ in Proc. IEEE Int. Conf. Edge Comput. (EDGE), Jun.2017, pp.232–235.
[0122] [4] S. Ranadheera, S. Maghsudi, and E. Hossain, ‘‘Computation offloading and activation of mobile edge computing servers: A minority game,’’ 2017, arXiv:1710.05499. [Online]. Available: http: / / arxiv.org / abs / 1710.05499.
[0123] [5] D. Mazza, D. Tarchi, and G. E. Corazza, ‘‘A cluster-based computation offloading technique for mobile cloud computing in smart cities,’’ in Proc. IEEE Int. Conf. Commun. (ICC), May 2016, pp.1–6.
Claims
CLAIMS 1. A method (200) for offloading sub-modules to edge nodes of an edge network, the method comprising: constructing (s202) a feasible sub-module offload pattern, FSOP, matrix for a set of offload sub-modules; choosing (s204) a set of one or more offload sub-modules to be assigned to a specific edge node for offloading; updating (s206) a sub-module offload pattern, SOP, matrix based on the FSOP matrix and the set of offload sub-modules chosen to be assigned to the edge node for offloading; and offloading (s208) the set of offload sub-modules in accordance with the SOP matrix.
2. The method of claim 1, wherein the FSOP matrix is of order Lnx J, Lnis the total number of sub-modules in the set of offload sub-modules, and J is the total number of feasible sub-module offload patterns.
3. The method of claim 2, wherein J =12 � ^^^^^^^^2+ ^^^^^^^^� + 1.
4. The method of any one of claims 1-3, wherein each row of the FSOP matrix correspond to a different one of the offload sub-modules, and each column of the FSOP matrix identifies a feasible sub-module offloading pattern.
5. The method of claim 4, wherein each column of the FSOP matrix identifies either a single offload sub-module or a set of two or more adjacent offload sub-modules.
6. The method of any one of claims 1-5, wherein choosing a set of one or more offload sub-modules to be assigned to a specific edge node for offloading comprises choosing theoffload sub-modules such that a quality of service, QoS, requirement will be satisfied while at the same time a weighted sum of a system execution time and energy consumption is minimized.
7. The method of any one of claims 1-6, wherein updating the SOP matrix based on the FSOP matrix and the set of offload sub-modules chosen to be assigned to the edge node for offloading comprising: determining a column of the FSOP matrix that identifies the set of offload sub-modules chosen to be assigned to the edge node for offloading; and setting element m*,c of the SOP matrix to 1, wherein m* is the index for the edge node, and c is the column number of the determined column of the FSOP matrix.
8. The method of claim 7, further comprising: prior to updating the SOP matrix, determining the set of offload sub-modules chosen to be assigned to the edge node for offloading.
9. The method of claim 8, wherein the set of offload sub-modules chosen to be assigned to the edge node for offloading is determined based on a current state of the SOP matrix and a set of offload sub-modules most recently choses to be assigned to the edge node for offloading.
10. The method of any one of claims 7-9, wherein the FSOP matrix is of order Lnx J, where Lnis the total number of sub-modules in the set of offload sub-modules, and J is the total number of feasible sub-module offload patterns, the SOP matrix is of order M x J, were M is a number of available edge nodes, and updating the SOP matrix further comprises, prior to setting element m*,c of the SOP matrix to 1, setting elements m*,1 to m*,J to 0.
11. A computing apparatus (500) for offloading sub-modules to edge nodes of an edge network, the computing apparatus being configured to perform a method (200) comprising: constructing a feasible sub-module offload pattern, FSOP, matrix for a set of offload sub- modules;choosing a set of one or more offload sub-modules to be assigned to a specific edge node for offloading; updating a sub-module offload pattern, SOP, matrix based on the FSOP matrix and the set of offload sub-modules chosen to be assigned to the edge node for offloading; and offloading the set of offload sub-modules in accordance with the SOP matrix.
12. The computer apparatus of claim 11, wherein the FSOP matrix is of order Lnx J, Lnis the total number of sub-modules in the set of offload sub-modules, and J is the total number of feasible sub-module offload patterns.
13. The computer apparatus of claim 12, wherein J =1 22 � ^^^^^^^^+ ^^^^^^^^� + 1.
14. The computer apparatus of any one of claims 11-13, wherein each row of the FSOP matrix correspond to a different one of the offload sub-modules, and each column of the FSOP matrix identifies a feasible sub-module offloading pattern.
15. The computer apparatus of claim 14, wherein each column of the FSOP matrix identifies either a single offload sub-module or a set of two or more adjacent offload sub-modules.
16. The computer apparatus of any one of claims 11-15, wherein choosing a set of one or more offload sub-modules to be assigned to a specific edge node for offloading comprises choosing the offload sub-modules such that a quality of service, QoS, requirement will be satisfied while at the same time a weighted sum of a system execution time and energy consumption is minimized.
17. The computer apparatus of any one of claims 11-16, wherein updating the SOP matrix based on the FSOP matrix and the set of offload sub-modules chosen to be assigned to the edge node for offloading comprising:determining a column of the FSOP matrix that identifies the set of offload sub-modules chosen to be assigned to the edge node for offloading; and setting element m*,c of the SOP matrix to 1, wherein m* is the index for the edge node, and c is the column number of the determined column of the FSOP matrix.
18. The computer apparatus of claim 17, wherein the method further comprises: prior to updating the SOP matrix, determining the set of offload sub-modules chosen to be assigned to the edge node for offloading.
19. The computer apparatus of claim 18, wherein the set of offload sub-modules chosen to be assigned to the edge node for offloading is determined based on a current state of the SOP matrix and a set of offload sub-modules most recently choses to be assigned to the edge node for offloading.
20. The computer apparatus of any one of claims 17-19, wherein the FSOP matrix is of order Lnx J, where Lnis the total number of sub-modules in the set of offload sub-modules, and J is the total number of feasible sub-module offload patterns, the SOP matrix is of order M x J, were M is a number of available edge nodes, and updating the SOP matrix further comprises, prior to setting element m*,c of the SOP matrix to 1, setting elements m*,1 to m*,J to 0.
21. A computer program (543) comprising instructions (544) which when executed by processing circuitry (502) of a computing apparatus (500) causes the computing apparatus to perform the method of any one of claims 1-10.
22. A carrier containing the computer program of claim 21, wherein the carrier is one of an electronic signal, an optical signal, a radio signal, and a computer readable storage medium (542).
