Competitive Online Virtual Cluster Embedding Algorithms

TL;DR

This paper introduces and analyzes two competitive online algorithms and two heuristics for the virtual cluster embedding problem, focusing on maximizing profit while respecting resource constraints in datacenter networks.

Contribution

It presents the first competitive online algorithms with proven ratios for the virtual cluster embedding problem, along with heuristic methods and comprehensive evaluation.

Findings

Different algorithms perform best under different request patterns.

Competitive algorithms guarantee worst-case performance bounds.

Heuristics effectively respect capacity constraints without competitive guarantees.

Abstract

In the conventional cloud service model, computing resources are allocated for tenants on a pay-per-use basis. However, the performance of applications that communicate inside this network is unpredictable because network resources are not guaranteed. To mitigate this issue, the virtual cluster (VC) model has been developed in which network and compute units are guaranteed. Thereon, many algorithms have been developed that are based on novel extensions of the VC model in order to solve the online virtual cluster embedding problem (VCE) with additional parameters. In the online VCE, the resource footprint is greedily minimized per request which is connected with maximizing the profit for the provider per request. However, this does not imply that a global maximization of the profit over the whole sequence of requests is guaranteed. In fact, these algorithms do not even provide a worst case guarantee on a fraction of the maximum achievable profit of a certain sequence of requests. Thus, these online algorithms do not provide a competitive ratio on the profit. In this thesis, two competitive online VCE algorithms and two heuristic algorithms are presented. The competitive online VCE algorithms have different competitive ratios on the objective function and the capacity constraints whereas the heuristic algorithms do not violate the capacity constraints. The worst case competitive ratios are analyzed. After that, the evaluation shows the advantages and disadvantages of these algorithms in several scenarios with different request patterns and profit metrics on the fat-tree and MDCube datacenter topologies. The results show that for different scenarios, different algorithms have the best performance with respect to certain metrics.