Section: New Results

Resource Allocation in Large Data Centres

Participants : Christine Fricker, Philippe Robert, Guilherme Thompson, Veronica Quintuna Rodriguez.

Efficient resource allocation in large data centers has become crucial matter since the expansion in volume and in variety of the internet based services and applications. Everyday examples, such as Video-on-Demand and Cloud Computing are part of this change in the internet environment, bringing new perspectives and challenges with it. Resource pooling (gathering resources to avoid idleness) and resource decentralization (to bring the service “closer" to the user) are too an important topic in service design, specially because of the inherent dichotomy presented in this discussion. Understanding and assessing the performance of such systems ought enable to better resource management and, consequently, better quality of service.

Currently, most systems operate under decentralized policies due to the complexity of managing data exchange on large scale. In such systems, customer demands are served respecting their initial service requirements (a certain video quality, amount of memory or processing power etc.) until the system reaches saturation, which then leads to the blockage of subsequent customer demands. Strategies that rely on the scheduling of tasks are often not suitable to address this load balancing problem as the users expect instantaneous service usage in real time applications, such as video transmission and elastic computation. Our research goal is to understand and redesign its algorithms in order to develop decentralized schemes that can improve global performance using local instantaneous information. This research is made in collaboration with Fabrice Guillemin, from Orange Labs.

In a first approach to this problem, we examined offloading schemes in fog computing context, where one data centers are installed at the edge of the network. We analyze the case with one data center close to user which is backed up by a central (usually bigger) data center. In [10], when a request arrives at an overloaded data center, it is forwarded to the other data center with a given probability, in order to help dealing with saturation and reducing the rejection of requests. In [17], we studied another scheme, where requests are systematically forwarded by the small data to a larger one, but with some trunk reservation to ensure service performance in the second one. We have been able to demonstrate the behavior and performance of these systems, using the invariant distribution of a random walks in the quarter plane, and obtaining explicit expressions for both schemes. Those two papers shed some light in the effectiveness of this fog computing design, by investigating two basic and intuitive policies, whose advantages can now be compared.

In [11] and [16], we investigated allocation schemes which consist in reducing the bandwidth of arriving requests to a minimal value. In the first, this process is initiated when the system is saturated and in the second when the system is close to saturation. We analyzed the effectiveness of such a downgrading policies. In the case of downgrading at saturation, we were able to find an explicit expression of the key performance metrics when two types of customers share a resource and type two asks for the double of resources compared to type one. And, for the second case, we could show that if the system is correctly designed then we can stop losing clients. We developed a mathematical model which allows us to predict system behavior under such a policy and calculate the optimal threshold (in the same scale as the resource) after which downgrading should be initiated. We proved the existence of a unique equilibrium point, around which we have been able to determine the probability a customer receives service at requested quality. We have also shown that system blockage becomes indeed negligible. This policy finds a natural application in the framework of video streaming services and other real time applications. Notably, we are able to derive explicit and simple expressions for many aspects of this system, giving special predictability the outcome of such policy.

Recently, we started to investigate the framework of network function virtualization, another emergent stream stream of research in resource allocation. We start by considering the execution of Virtualized Network Functions (VNFs) in data centers whose capacities are limited and service execution time is constrained by telecommunication protocols. Virtualization practices play a crucial role in the evolution of telecommunications network archi­tectures, since the service providers can reduce the investment on the edge and share resource more efficiently. Macrofunctions are virtuzatized into micro ones and treat individually. Through simulations and basic mathematical models, we aroused the discussion of three different prioritization policies and their trade-offs. The have shown that in for parallelizable macrofunctions (i.e. no order of execution), the greedy algorithm ensures the best performance in terms of execution delay. For chained ones, macrofunctions whose microfunctions need to be run in a certain order, this algorithm is not suitable, the Round Robin and the Dedicated Core policies perform with the same level.

With these results in mind, we have extend our research towards more complex systems, investigating the behaviour of multiple resource systems (such as a Cloud environment, where computational power is provided using unities of CPU and GB of RAM). We analyzed cooperation between data centers offering multiple resources and under imbalanced loads, a problem that naturally arises from the decentralization of resources. Again, we consider instantaneous service. By forwarding some clients across the system, we could design a policy that is allows cooperation between system and preserves service quality at both data centers. We consider two types of demands asking for two types of resources; particularly, type one clients demand more of type one resource (and symmetrically for type two). We have shown that under our forwarding scheme, which offloads clients requiring most of the saturated resource locally at each data center, we can eliminate losses (in a well design system). Some other interesting properties that can help systems designers are as well derived, such as the minimum threshold for the sustainability of such scheme and the offloading rates. A document is being written to further publication.