A scalable model for multi-period virtual network mapping for resilient multi-site data centers

Abstract

Abstract-In the currently dominant cloud computing paradigm, applications are being served in data centers (DCs) which are connected to high capacity optical networks. For cost efficiency reasons, in both DC and optical network domains, virtualization of the physical hardware is exploited. In a DC, it means that multiple so-called virtual machines (VMs) are being hosted on the same physical server. Similarly, the network is partitioned into separate virtual networks, thus providing isolation between distinct virtual network operators (VNOs). Thus, the problem of virtual network mapping arises: how to decide which physical resources to allocate for a particular virtual network? In this paper, we study that problem in the context of cloud computing with multiple DC sites. This introduces additional flexibility, due to the anycast routing principle: we have the freedom to decide at what particular DC location to serve a particular application. We can exploit this choice to minimize the required resources when solving the virtual network mapping problem. This paper builds on our earlier work and solves the resilient virtual network mapping problem that optimally decides on the mapping of both network and data center resources, considering timevarying traffic conditions and protecting against possible failures of both network and DC resources. Previously, we developed a model to solve the multi-period traffic case one step at a time: given the virtual network mapping in period t, we determine the (possibly changed) mapping for t + 1. Compared to that previous work, we now (i) define a truly multi-period model path formulation exploiting column generation, and (ii) demonstrate its scalability on a nation-wide network with traffic that varies across multiple periods.

Exploring the Benefit of Rerouting Multi-Period Traffic to Multi-Site Data Centers [Invited]

“…Note that our model assumes that we can aggregate all traffic originating at the same source: we do not individually model unit requests that share the same source and holding time. This improves scalability of the model compared to, e.g., our earlier work [5].…”

“…Specifically, we consider the case where routing, and thus also DC selection, can be revised at discrete points in time: we assume the volume of service requests to vary over time, which we assume to be divided in multiple periods. Compared to our previous work in this area, i.e., [5], we provide the following contributions:…”

“…Since then, we have developed optimization models to jointly optimize the routing (in terms of both primary and shared backup paths) over multiple periods together. As stated before, we reported initial results in the short conference papers [5], [2]. Next, we will introduce the exact problem statement and then disclose the full model details before reporting on experiments.…”

Anycast (re)routing of multi-period traffic in dimensioning resilient backbone networks for multi-site data centers

“…The RMP's task is to find the optimal combination out of a set of candidate "configurations", C, while the PP's task is to find/create new configuration(s) that, when added to the pool C for the RMP leads to a solution with a better objective value. More concretely, a "configuration" in our model is associated with a given source node v s , and comprises a set of three paths as sketched in Fig Compared to [3], our new model improves scalability by (i) aggregating all unit demands from a single source node and with the same holding time into a single request, and (ii) check for reconfigurations (which we want to minimize, as a secondary objective next to the total amount of network resources) based on paths rather than links. For details, we refer to [4, Section 4.4].…”

“…We build on our latest work on this topic [3], and contribute with (i) a new column generation model that is path-based (rather than link-based), (ii) an exploration of the effect of parallel execution of multiple so-called pricing problems (PPs), and (iii) more extensive experiments that also consider variations in the choice of DC locations.…”