Ferhat Dikbiyik

Design of Disaster-Resilient Optical Datacenter Networks

Survivability against disasters-both natural and deliberate attacks, and spanning large geographical areas-is becoming a major challenge in communication networks. Cloud services delivered by datacenter networks yield new opportunities to provide protection against disasters. Cloud services require a network substrate with high capacity, low latency, high availability, and low cost, which can be delivered by optical networks. In such networks, path protection against network failures is generally ensured by providing a backup path to the same destination (i.e., a datacenter), which is link-disjoint to the primary path. This protection fails to protect against disasters covering an area which disrupts both primary and backup paths. Also, protection against destination (datacenter) node failure is not ensured by a generic protection scheme. Moreover, content/service protection is a fundamental problem in a datacenter network, as the failure of a datacenter should not cause the disappearance of a specific content/service from the network. So content placement, routing, and protection of paths and content should be addressed together. In this work, we propose an integrated Integer Linear Program (ILP) to design an optical datacenter network, which solves the above-mentioned problems simultaneously. We show that our disaster protection scheme exploiting anycasting provides more protection, but uses less capacity than dedicated single-link failure protection. We show that a reasonable number of datacenters and selective content replicas with intelligent network design can provide survivability to disasters while supporting user demands. We also propose ILP relaxations and heuristics to solve the problem for large networks.

Network adaptability from disaster disruptions and cascading failures

Disasters can cause severe service disruptions due to large-scale correlated cascading failures in telecom networks. Major network disruptions due to disasters - both natural (e.g., Hurricane Sandy, 2011 Japan Tsunami) and human-made (e.g., 9/11 terrorist attack) - deprive the affected population of essential network services for weeks and severely hamper rescue operations. Many techniques exist to provide fast network protection, but they are optimized for limited faults without addressing the extent of disaster failures. Thus, there is a pressing need for novel robust survivability methods to mitigate the effects of disasters on telecom networks. While researchers in climatology, geology, and environmental science have been studying how to predict disasters and assess disaster risks for certain regions, networking research can exploit this information to develop novel methods to prepare networks to handle disasters with the knowledge of risky regions and to better prepare them for a predicted disaster. The events during the aftermath of a disaster should also be considered. For instance, methods to re-arrange network resources and services on a partially damaged network, which is the property of a self-organizing network, should be developed, and new algorithms to manage the post-disaster traffic deluge and to relieve the rescue operations after a disaster, with the knowledge of the post-disaster failures, should be investigated. Since cloud services today are an integral part of our society and massive amounts of content/services have been created and shared over the cloud, loss/disruption of critical content/ services caused by disasters can significantly affect the security and economic well being of our society. As the network is becoming increasingly an end-to-content (vs. end-to-end) connection provider, we have to ensure reachability of content from any point of a network, which we call content connectivity (in contrast to network connectivity) after disaster failures. This article presents the nature of possible disruptions in telecom networks caused by disaster events, and sheds light on how to prepare the network and cloud services against disasters, and adapt them for disaster disruptions and cascading failures.

Minimizing the Risk From Disaster Failures in Optical Backbone Networks

Failures caused by disasters (e.g., weapons of mass destruction (WMD) attacks, earthquakes, hurricanes, etc.) can create huge amount of loss and disruptions in optical backbone networks. Due to major network disruptions in recent disasters, network operators need solutions to prevent connections from disasters, and recover them after disasters. 1) Prevention requires proactive approaches where the damage from a possible disaster should be estimated. Specifically, we propose disaster-risk-aware provisioning, which minimizes loss to a network operator in case of a disaster. 2) Recovery methods should consider that, after the initial failure, more connections might be disconnected by correlated cascading failures. Thus, we investigate a reprovisioning scheme to recover disrupted connections. Numerical examples conducted for different disaster types (WMD attack, earthquake, and tornado) show that our schemes significantly reduce the risk and loss in case of a disaster.

Disaster-aware datacenter placement and dynamic content management in cloud networks

Recent targeted attacks and natural disasters have made disaster-resilient cloud network design an important issue. Network operators are investigating proactive and reactive measures to prevent huge data loss and service disruptions in case of a disaster. We present novel techniques for disaster-aware datacenter placement and content management in cloud networks that can mitigate such loss by avoiding placement in given disaster-vulnerable locations. We first solve a static disaster-aware datacenter and content placement problem by adopting an integer linear program with the objective to minimize risk, defined as expected loss of content. It is a measure of how much, in terms of cost or penalty, a network operator may lose probabilistically due to possible disasters in a cloud network. We also show how a service providers budget constraint can affect disaster-aware placement design. Since disaster scenarios, content popularity, and/or importance are always changing in time, content placement should rapidly adapt to these changes. We propose a disaster-aware dynamic content-management algorithm that can adjust the existing placement based on dynamic settings. Besides reducing the overall risk and making the network disaster-aware, reducing network resource usage and satisfying quality-of-service requirements can also be achieved in this approach. We also provide a cost analysis of employing a dynamic disaster-aware placement design in the network based on real-world cloud pricing.

Network adaptability to disaster disruptions by exploiting degraded-service tolerance

The rapid increase in network traffic with new bandwidth-hungry applications such as cloud computing and telemedicine makes disaster survivability a crucial concern as the data (and revenue) loss caused by large-scale correlated cascading failures can be very high. To alleviate their impact, new measures should be taken since the nature of the network changes dramatically as available resources decrease during disasters. We develop a metric, called degraded-service tolerance, which can reduce protection cost and network disruption, and support maximal carried traffic in case of disasters. Degraded-service-tolerant connections can be admitted and recovered with reduced bandwidth under resource crunch. Our scheme re-assigns resources among connections by leveraging their degraded-service tolerance. A case study shows how our proposal can be applied to boost network performance during the resource crunch following a disaster.

Exploiting excess capacity to improve robustness of WDM mesh networks

Excess capacity (EC) is the unused capacity in a network. We propose EC management techniques to improve network performance. Our techniques exploit the EC in two ways. First, a connection preprovisioning algorithm is used to reduce the connection setup time. Second, whenever possible, we use protection schemes that have higher availability and shorter protection switching time. Specifically, depending on the amount of EC available in the network, our proposed EC management techniques dynamically migrate connections between high-availability, high-backup-capacity protection schemes and low-availability, low-backup-capacity protection schemes. Thus, multiple protection schemes can coexist in the network. The four EC management techniques studied in this paper differ in two respects: when the connections are migrated from one protection scheme to another, and which connections are migrated. Specifically, Lazy techniques migrate connections only when necessary, whereas Proactive techniques migrate connections to free up capacity in advance. Partial Backup Reprovisioning (PBR) techniques try to migrate a minimal set of connections, whereas Global Backup Reprovisioning (GBR) techniques migrate all connections. We develop integer linear program (ILP) formulations and heuristic algorithms for the EC management techniques. We then present numerical examples to illustrate how the EC management techniques improve network performance by exploiting the EC in wavelength-division-multiplexing (WDM) mesh networks.

A disaster-resilient multi-content optical datacenter network architecture

Cloud services based on datacenter networks are becoming very important. Optical networks are well suited to meet the demands set by the high volume of traffic between datacenters, given their high bandwidth and low-latency characteristics. In such networks, path protection against network failures is generally ensured by providing a backup path to the same destination, which is link-disjoint to the primary path. This protection fails to protect against disasters covering an area which disrupts both primary and backup resources. Also, content/service protection is a fundamental problem in datacenter networks, as the failure of a single datacenter should not cause the disappearance of a specific content/service from the network. Content placement, routing and protection of paths and content are closely related to one another, so the interaction among these should be studied together. In this work, we propose an integrated ILP formulation to design an optical datacenter network, which solves all the above-mentioned problems simultaneously. We show that our disaster protection scheme exploiting anycasting provides more protection, but uses less capacity, than dedicated single-link protection. We also show that a reasonable number of datacenters and selective content replicas with intelligent network design can provide survivability to disasters while supporting user demands.

Disaster-aware service provisioning by exploiting multipath routing with manycasting in telecom networks

Disasters may cause large-area failures in high-capacity telecom networks, leading to huge data loss. Survivable service provisioning is crucial to minimize the effects of network / datacenter failures and maintain critical services in case of a disaster. We propose a novel disaster-aware service-provisioning scheme that multiplexes service over multiple paths destined to multiple servers/datacenters with manycasting. Our scheme maintains some bandwidth (i.e., reduced service) after a disaster failure vs. no service at all. Numerical examples show that our approach offers high level of survivability against link and node failures that may be caused by disasters and post-disaster failures at no extra cost compared to the other survivable schemes.

Disaster-aware service provisioning with manycasting in cloud networks

Cloud services delivered by high-capacity optical datacenter networks are subject to disasters which may cause large-area failures, leading to huge data loss. Survivable service provisioning is crucial to minimize the effects of network/datacenter failures and maintain critical services in case of a disaster. We propose a novel disaster-aware service-provisioning scheme that multiplexes service over multiple paths destined to multiple servers/datacenters with manycasting. Our scheme maintains some bandwidth (i.e., degraded service) after a disaster failure vs. no service at all. We formulate this problem into a mathematical model which turns out to be an Integer Linear Program (ILP), and we provide heuristic optimization approaches as ILP is intractable for large problem instances. Numerical examples show that exploiting manycasting by intelligently selecting destinations in a risk-aware manner for service provisioning offers high level of survivability against link and node failures that may be caused by disasters and post-disaster failures at no extra cost compared to the other survivable schemes.

Disaster-aware submarine fiber-optic cable deployment

Network survivability is an important element in telecommunication network design nowadays because of the social and economic reliance on the Internet and the significant cost associated with service interruption. Moreover, the fact that submarine fiber-optic cables are susceptible to man-made or natural disasters, such as earthquakes, is well recognized. A disaster-resilient submarine cable deployment can save cost incurred by network operators such as the capacity loss cost, the cruising cost and the repair cost of the damaged cables, in order to restore network service when a cable break is prompted by a disaster occurrence. In this work, we investigate disaster-aware submarine cable deployment problem. While selecting a path for the cables, our approach aims to minimize the total expected loss cost, considering that submarine fiber-optic cables may break because of natural disasters, subject to deployment budget, path uniqueness, regular protection, elliptic shape, and linearization constraints. In our approach, we assume disaster-unrelated failures are handled by providing a backup cable along with primary cable. We consider a scenario with two nodes located on two different lands separated by a water body (sea/ocean). We then consider an elliptic cable shape to formulate the problem, which can be extended to other cable shapes, subject to avoiding deploying cable in disaster zones. We provide an Integer Linear Programming formulation for the problem. Finally, we present illustrative numerical examples that show the potential benefit of our approach as well as conclusion and future work.