Optimal bandwidth-aware VM allocation for Infrastructure-as-a-Service

TL;DR

This paper addresses the complex problem of optimally allocating virtual machines with bandwidth guarantees in data centers, proving its NP-hardness even in simplified topologies, and providing dynamic programming solutions for practical request sizes.

Contribution

It proves NP-hardness of bandwidth-aware VM allocation in tree networks and introduces efficient dynamic programming algorithms for small request graphs.

Findings

NP-hardness holds even for tree topologies

Dynamic programming algorithm runs in O(3^kn) for small requests

Modified algorithm achieves O(k^2n) for clique request topologies

Abstract

Infrastructure-as-a-Service (IaaS) providers need to offer richer services to be competitive while optimizing their resource usage to keep costs down. Richer service offerings include new resource request models involving bandwidth guarantees between virtual machines (VMs). Thus we consider the following problem: given a VM request graph (where nodes are VMs and edges represent virtual network connectivity between the VMs) and a real data center topology, find an allocation of VMs to servers that satisfies the bandwidth guarantees for every virtual network edge---which maps to a path in the physical network---and minimizes congestion of the network. Previous work has shown that for arbitrary networks and requests, finding the optimal embedding satisfying bandwidth requests is $\mathcal{NP}$-hard. However, in most data center architectures, the routing protocols employed are based on a spanning tree of the physical network. In this paper, we prove that the problem remains $\mathcal{NP}$-hard even when the physical network topology is restricted to be a tree, and the request graph topology is also restricted. We also present a dynamic programming algorithm for computing the optimal embedding in a tree network which runs in time $O(3^kn)$, where $n$ is the number of nodes in the physical topology and $k$ is the size of the request graph, which is well suited for practical requests which have small $k$. Such requests form a large class of web-service and enterprise workloads. Also, if we restrict the requests topology to a clique (all VMs connected to a virtual switch with uniform bandwidth requirements), we show that the dynamic programming algorithm can be modified to output the minimum congestion embedding in time $O(k^2n)$.