Multiuser Computation Offloading and Downloading for Edge Computing With Virtualization

Yuan Liu (刘元), Zezu Liang, Yuan Liu, Tat-Ming Lok, Kaibin HuangIEEE Trans. Wireless Communications 2019 · DOI

TL;DR

This paper tackles the challenge of optimizing computation offloading in multi-user mobile-edge computing (MEC) systems, specifically addressing the performance degradation caused by I/O interference between virtual machines (VMs) on edge servers. The authors propose novel algorithms for joint radio-and-computation resource allocation that maximize offloading throughput or minimize mobile energy consumption, outperforming existing methods by accounting for I/O interference and download latency.

What Problem Does This Paper Address?

Mobile-edge computing (MEC) offers a promising solution for latency-sensitive mobile applications by offloading computation tasks to nearby edge servers. However, a key challenge arises from the virtualization of these edge servers, where multiple virtual machines (VMs) share resources. This sharing leads to "I/O interference," which degrades the computational performance of individual VMs. Existing research has largely overlooked this interference, assuming a fixed computational capacity for servers. Furthermore, the communication overhead for downloading computation results, often assumed negligible, can be significant for applications like augmented reality and image processing. This paper addresses the critical need to jointly optimize radio and computation resource allocation in multiuser MEC systems, explicitly considering both I/O interference and result downloading latency.

Methods

The authors formulate two main optimization problems: one for maximizing the sum offloading rate and another for minimizing the sum mobile energy consumption. Both problems are framed as non-convex mixed-integer programming problems.

For sum offloading rate maximization, they employ a decomposition approach, breaking the problem into master and slave sub-problems. The slave problem, which optimizes user selection, data sizes, and time division, is solved using an iterative algorithm based on the Dinkelbach method. The master problem then finds the optimal number of offloading users through a search.

For sum energy minimization, the problem is transformed into an equivalent one that facilitates scheduling. They categorize users into disjoint subsets based on their offloading needs and energy efficiency. A suboptimal algorithm is then developed using a set of rules to ensure latency constraints are met while minimizing energy consumption.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • I/O Interference is a Significant Bottleneck: The paper strongly emphasizes that neglecting I/O interference in virtualized MEC systems leads to unrealistic assumptions about server capacity and can significantly degrade performance.
  • Joint Optimization is Crucial: Effectively managing both radio resources (offloading/downloading) and computational resources (parallel processing on the edge) is essential for optimal system performance.
  • Result Downloading Matters: The communication overhead for downloading results cannot be ignored, especially for data-intensive applications, and must be incorporated into scheduling decisions.
  • Trade-offs Exist: There's a fundamental trade-off between offloading more data (higher throughput) and managing interference, as well as between offloading and local computation for energy savings.
  • Greedy Approaches Can Be Effective: For certain scenarios, greedy strategies based on user transmission rates or energy efficiency can yield near-optimal results.

Key Results

The proposed algorithms demonstrate superior performance compared to benchmark algorithms in terms of both sum offloading rate and sum energy consumption.

  • Offloading Rate: Simulations show that the proposed optimal algorithm achieves higher sum offloading rates, especially as the number of users increases. It also exhibits better resilience to I/O interference (degradation factor 'd') compared to other methods. For instance, when K=12, the optimal algorithm shows significant improvements over benchmarks.
  • Energy Minimization: The suboptimal algorithm for energy minimization also outperforms the all-offloading scheme, particularly when latency constraints are tight. For example, with T=30ms, the suboptimal algorithm achieves about 14% energy savings. The proposed algorithms also show better performance against increasing degradation factors.

Limitations and Future Work

  • Complexity: While the proposed algorithms are designed to be efficient, solving the optimization problems can still be computationally intensive, especially for a large number of users. The paper mentions that the complexity of the overall algorithm for rate maximization is O(K log K).
  • Suboptimal Solution for Energy Minimization: The algorithm for energy minimization is suboptimal, although it achieves good performance.
  • Future Work: The paper implicitly suggests future work could explore more advanced interference modeling, dynamic resource allocation in more complex network topologies (e.g., multi-cell MEC), and the integration of other factors like user mobility and diverse application requirements. The authors also highlight the need for further investigation into the trade-offs between different performance metrics.

Summary updated

BibTeX
@article{2019multiuser,
  author = {Yuan Liu (刘元) and Zezu Liang and Yuan Liu and Tat-Ming Lok and Kaibin Huang},
  title = {{Multiuser Computation Offloading and Downloading for Edge Computing With Virtualization}},
  journal = {IEEE Trans. Wireless Communications},
  year = {2019},
  doi = {10.1109/twc.2019.2922613},
}
Figure