2020 Fiscal Year Research-status Report
Mixed-Clairvoyance Task Offloading and Scheduling in Multi-access Edge Computing Systems: From Combinatorial Optimization to Machine Learning
| Project/Area Number |
20K19794
|
|---|
| Research Institution | Osaka Prefecture University |
|---|
Principal Investigator |
江 易翰 大阪府立大学, 工学(系)研究科(研究院), 助教 (10824196)
|
|---|
| Project Period (FY) |
2020-04-01 – 2023-03-31
|
| Keywords | Task offloading / Task scheduling / Deep learning / Internet of Things / Edge computing / Nonlinear programming |
|---|
| Outline of Annual Research Achievements |
(1) We investigate the problem of cotask processing in multi-access edge computing (MEC) systems, which can be characterized as a nonlinear program (NLP) to minimize total cotask completion time (TCCT).
(2) Due to the lack of the probability distribution of link data rates, we apply transformation techniques to render the NLP a more tractable problem.
(3) We design a deep learning method to make cotask offloading and scheduling decisions based on random perturbation.
(4) Our simulation results show the convergence and the TCCT performance of the proposed solution under various system settings.
|
| Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
In this fiscal year, what we have achieved can be summarized as follows.
(1) We formulated the problem of cotask processing in MEC systems as an NLP and show its NP-hardness.
(2) We applied a parameterization and a Lagrangian technique to transform the NLP to a parameterized and unconstrained version.
(3) We proposed the deep dual learning (DDL) method to update the primal and dual variables iteratively, where the learning parameters are governed by two deep neural networks (DNNs).
(4) We provided the duality gap and time complexity analyses to show the effectiveness of the DDL method.
(5) Our simulation results demonstrated that the DDL method can gradually converge over iterations and outperform other comparison schemes in terms of TCCT and the variance of CCT.
|
| Strategy for Future Research Activity |
In the next fiscal year, our goals can be enumerated as follows.
(1) We aim to address the problem of information sampling and transmission scheduling for serverless functions in the Internet of Things (IoT) systems with serverless computing to optimize the age of service (AoS) of serverless functions.
(2) We will leverage combinatorial optimization techniques to mathematically formulate the problem and investigate the problem intractability.
(3) We plan to design both offline and online age-efficient algorithms for the information sampling and transmission scheduling with and without the prior knowledge of the invocations of serverless functions, respectively.
(4) To evaluate the AoS performance, we will conduct simulations and testbed experiments to demonstrate that the proposed solution outperforms existing ones under various system parameters.
|
| Causes of Carryover |
In the next stage of this research, we plan to purchase a rack server equipped with higher computing powers for conducting large-scale simulations. Since the residual amount at the end of this fiscal year is insufficient for purchasing the equipment, we would like to use it in the next fiscal year alternatively.
|
